Deep learning-based segmentation, grading and genotype classification of glioma: a literature analysis Authors: Weihan Song, Yuan Hu, Yi Wang, Jian Ruan, Bo Lin, Weibin Yang Published: 2026 ABSTRACT Background and ObjectiveGlioma represents the most prevalent and lethal primary brain tumor in adults, and its high mortality poses a significant challenge to clinical decision-making. While the conventional gold standard for diagnosis relies on the examination of surgically resected tumor tissue, recent advances in artificial intelligence, particularly deep learning, have demonstrated compelling performance in medical image analysis, offering the potential for non-invasive prediction of the genetic and histological features of glioma as well as automatic tumor segmentation. This review aims to comprehensively summarize deep learning-based techniques for glioma segmentation, grading, and genotype classification.MethodsA systematic literature search was conducted across PubMed, Institute of Electrical and Electronics Engineers (IEEE) Xplore, and Scopus for articles published in English between January 2014 and December 2025, using combinations of the terms “glioma”, “brain tumor”, “deep learning”, “convolutional neural network”, “segmentation”, “classification”, “grading”, “genotype”, “isocitrate dehydrogenase (IDH)”, “1p/19q”, and “O6-methylguanine-DNA methyltransferase (MGMT)”. Peer-reviewed studies that applied deep learning methods to glioma analysis using medical imaging data and reported quantitative performance results were included, whereas review articles without original results, studies using exclusively non-deep-learning methods, and studies focused solely on non-glioma brain tumors were excluded.Key Content and FindingsThis review first summarizes the commonly used public datasets [Brain Tumor Segmentation (BraTS), The Cancer Imaging Archive (TCIA), The Cancer Genome Atlas (TCGA), synthetic, and private cohorts] and performance metrics, and then surveys deep learning-based approaches for glioma segmentation, grading, and genotype classification according to task. Recent methodological advances, including Vision Transformers, foundation models such as the segment anything model (SAM) and MedSAM, diffusion models, and self-supervised learning, are also discussed. Reported performance varies considerably across datasets and evaluation protocols, and risks of bias, including slice-level data leakage, patient overlap across public repositories, and single-center confounding, remain prevalent in the genotype prediction literature.ConclusionsDeep learning has become a transformative tool for non-invasive glioma characterization. Future research should prioritize prospective multi-center validation, standardized evaluation protocols, calibrated uncertainty estimation, and clinically deployable frameworks, thereby facilitating the translation of these techniques into routine diagnostic workflows and ultimately improving outcomes for patients with glioma. Introduction Gliomas, which originate from glial cells, have become the most prevalent category of primary brain tumors. The World Health Organization (WHO) historically classified gliomas into four grades (I-IV) based on microscopic characteristics and tumor behavior (1). Grades I and II were classified as low-grade gliomas (LGGs), characterized by relatively benign nature and slow growth, whereas grades III and IV represented high-grade gliomas (HGGs), distinguished by aggressive and invasive attributes. In 2021, the fifth edition of the WHO Classification of Central Nervous System Tumors (WHO-CNS5) introduced an integrated diagnostic framework for adult-type diffuse gliomas that prioritizes molecular markers over histology alone (1). Under this framework, adult-type diffuse gliomas are stratified into three principal entities: astrocytoma, isocitrate dehydrogenase (IDH)-mutant; oligodendroglioma, IDH-mutant and 1p/19q-codeleted; and glioblastoma, IDH-wildtype. Histological grade is handled within these molecular categories – for instance, astrocytoma, IDH-mutant, is assigned WHO-CNS grades 2, 3, or 4, whereas oligodendroglioma, IDH-mutant and 1p/19q-codeleted, is assigned grades 2 or 3. Glioblastoma, IDH-wildtype, corresponds to WHO-CNS grade 4 by definition. Additional molecular markers, including O6-methylguanine-DNA-methyltransferase (MGMT) promoter methylation status, provide further prognostic and predictive information that influences treatment planning. The diagnosis of glioma has historically entailed the surgical removal of affected tissue from patients. However, craniotomy procedures entail inherent high risks and uncertainties, and the subsequent waiting period for pathological detection results is deemed inefficient. Over time, an array of imaging techniques, including positron emission tomography (PET), computed tomography (CT), and magnetic resonance imaging (MRI), has been employed for brain tumor screening. Medical professionals utilize the information obtained from these imaging modalities to extract essential details about gliomas and provide preliminary estimates (2). Notably, MRI has emerged as the preeminent modality for both pre- and post-surgical brain tumor diagnosis, due to its exceptional performance (3). The diverse modalities of MRI, including diffusion-weighted imaging (DWI), T1-weighted, T2-weighted, T1-weighted contrast-enhanced (T1CE), and fluid-attenuated inversion recovery (FLAIR) MRI, capture various aspects of gliomas, providing distinct perspectives and features that facilitate more precise diagnoses (4). Although advanced MRI has substantially reduced the reliance on purely invasive procedures for initial assessment, imaging-based evaluation remains an adjunct to, rather than a replacement for, tissue-based diagnosis and molecular testing. The interpretation of images still necessitates a considerable investment of time from experienced clinicians. Furthermore, discrepancies in interpretations among different medical practitioners and the inherent subjectivity of individual assessments must be acknowledged. The evolution of computer science has led to the widespread adoption of medical imaging analysis and computer-aided diagnosis technology in clinical settings. These advancements furnish clinicians with valuable information, aiding in accurate early diagnosis. Notably, in the field of medical image analysis, considerable attention has been directed toward glioma segmentation, grading, and genotype classification, yielding stable progress and promising performance in research endeavors. The tasks of glioma segmentation, grading, and genotype classification involve a structured process (5). Firstly, the system receives a designated input image. Subsequently, the system automatically segments the tumor area (6) from normal tissues by employing feature extraction and analysis, which includes considerations such as location, texture and shape (7). Lastly, the system produces outputs comprising a segmentation map, the tumor grade, and genotype information corresponding to the input image. In contrast to conventional methodologies, the employment of deep learning methods significantly enhances the efficiency and accuracy of the diagnostic process, concurrently reducing time consumption and overall costs (8). In the future, the application of deep learning methodologies is anticipated to broaden its scope in glioma segmentation, grading, and genotype classification (9). Meanwhile, the diagnostic framework for glioma has shifted from purely histological assessment toward integrated molecular classification under WHO-CNS5, and methodological advances such as Vision Transformers, foundation models, diffusion models, and self-supervised learning (SSL) have rapidly reshaped the research landscape. However, existing surveys have not comprehensively covered these recent developments across all three tasks within a unified framework, nor have they critically appraised the evaluation protocols and risks of bias that undermine the reliability of reported performance, particularly in the genotype prediction literature. An up-to-date and methodologically rigorous synthesis is therefore warranted. In this paper, we present a comprehensive review of deep learning-based techniques for glioma segmentation, grading, and genotype classification. The remainder of this article is organized as follows: we first describe the literature search methodology, followed by the background and preliminaries, including a taxonomy of methods and a comparison with prior surveys. We then introduce the commonly used datasets and summarize the evaluation metrics together with recent architectural advances. Next, deep learning-based glioma segmentation, grading, and genotype classification methods are reviewed in turn. Finally, we discuss the existing common challenges, outline directions for future research, and conclude the survey. We present this article in accordance with the Narrative Review reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0730/rc). Methods A systematic literature search was conducted across three major electronic databases—PubMed, Institute of Electrical and Electronics Engineers (IEEE) Xplore, and Scopus—to identify relevant studies. The search strategy employed combinations of the following keywords: “glioma”, “brain tumor”, “deep learning”, “convolutional neural network”, “segmentation”, “classification”, “grading”, “genotype”, “IDH”, “1p/19q”, and “MGMT.” The search was restricted to articles published in English between January 2014 and December 2025. Inclusion criteria were: (I) peer-reviewed journal articles or conference proceedings; (II) studies that applied deep learning methods to glioma analysis using medical imaging data; and (III) studies that reported quantitative performance results. Exclusion criteria were: (I) review articles without original experimental results; (II) studies using exclusively non-deep-learning methods; and (III) studies focused solely on non-glioma brain tumors. Two authors (W.S. and Y.H.) independently screened the titles and abstracts of all retrieved records and assessed the full texts of potentially eligible articles, and any disagreements were resolved by discussion with a third author (W.Y.) until consensus was reached. The reference lists of included studies and pertinent reviews were also manually screened to identify additional eligible articles. The search strategy is summarized in Table 1, which also presents the detailed search strategy of PubMed as an example. Table 1The search strategy summaryItemsSpecificationDate of searchDecember 31, 2025Databases and other sources searchedPubMed, IEEE Xplore, and Scopus; the reference lists of included studies and pertinent reviews were also screened manuallySearch terms usedCombinations of “glioma”, “brain tumor”, “deep learning”, “convolutional neural network”, “segmentation”, “classification”, “grading”, “genotype”, “IDH”, “1p/19q”, and “MGMT”. Detailed PubMed strategy as an example: (“glioma”[MeSH Terms] OR “glioma”[Title/Abstract] OR “brain tumor”[Title/Abstract]) AND (“deep learning”[Title/Abstract] OR “convolutional neural network”[Title/Abstract]) AND (“segmentation”[Title/Abstract] OR “classification”[Title/Abstract] OR “grading”[Title/Abstract] OR “genotype”[Title/Abstract] OR “IDH”[Title/Abstract] OR “1p/19q”[Title/Abstract] OR “MGMT”[Title/Abstract]); filters: English language, publication date from 2014/01/01 to 2025/12/31TimeframeJanuary 1, 2014 to December 31, 2025Inclusion and exclusion criteriaInclusion: peer-reviewed journal articles or conference proceedings that applied deep learning methods to glioma analysis using medical imaging data and reported quantitative performance results; English-language publications. Exclusion: review articles without original experimental results; studies using exclusively non-deep-learning methods; studies focused solely on non-glioma brain tumorsSelection processTwo authors (W.S. and Y.H.) independently screened titles and abstracts and assessed full texts of potentially eligible articles; disagreements were resolved by discussion with a third author (W.Y.) until consensus was reached The search strategy summary A total of 193 studies that met the eligibility criteria were ultimately included in this survey. No language restrictions other than English were applied, and no study design or sample size restrictions were imposed. Because this is a narrative review of previously published studies and publicly available de-identified datasets, no new patient data were collected, and institutional ethics approval and informed consent were not required. Background and preliminaries The deep neural network (DNN) emulates the cognitive processes of the human brain during data processing and pattern synthesis for decision-making (10). It comprises various specialized layers, including convolution, pooling, padding, normalization, among others. The construction of a comprehensive DNN involves the sequential stacking of these layers. In contrast to traditional machine learning approaches, deep learning-based methods exhibit the inherent capability to automatically extract high-dimensional discriminative features in an end-to-end manner, making them the most popular methods in medical image processing. LeNet was a successful application of convolutional neural network (CNN) in the 1990s, featuring two convolution layers, two pooling layers, and two fully connected layers, employing 5×5 convolution kernels with a stride of 1, and max-pooling in the pooling layers (11). Subsequently, with the availability of advanced graphic processing units (GPUs) and computing resources, AlexNet presented a deeper architecture comprising five convolutional layers and three fully connected layers (12). Since then, the evolution of CNN has seen the incorporation of various layers such as convolutional, nonlinearity, pooling, and fully connected layers. The refinement of CNN encompassed regularization, optimization, normalization, loss functions, and parameter initializations. For example, GoogleNet introduced the inception module to optimize network parameters with its different size of kernels (10). DenseNet innovatively added a minimal number of feature maps to the network, facilitating effective training of deeper network (13). ResNet addressed the challenge of gradient vanishing and explosion through the incorporation of shortcut edges and residual connections between convolution layers (14). By replacing the conventional fully connected layer of a typical CNN with a convolutional layer and incorporating a basic upsampling strategy for pixel-wise labeling, the fully convolutional network (FCN) has emerged as a leading methodology in semantic segmentation (15). Inspired by FCN, the U-Net architecture was introduced and demonstrated outstanding performance, particularly in the domain of medical image processing (16). Over time, U-Net has gained prominence in medical image segmentation, with modifications to its loss function and structural adaptations being explored for further enhancements. Numerous derivatives of U-Net have been developed by researchers, including 3D U-Net (17), U-Net++ (18), V-Net (19), and nn U-Net (20). As shown in Figure 1, for using deep learning for segmentation, grading and genotype classification of gliomas a basic process is available today, but the design of an effective DNN entails a number of considerations. First and foremost, it is of paramount importance to procure a high-quality dataset that is exclusively dedicated to glioma cases. And the dataset must be adequately annotated to ensure an accurate training process. Secondly, the selection of an appropriate architectural design will facilitate the undertaking of the task. For example, the U-Net and its variants are frequently regarded as optimal choices for glioma segmentation. Third, the consideration of multi-modality of input images, attention mechanisms, transfer learning skills, regularization techniques, and optimization of loss functions are useful for the refinement of the process (21,22). Overview of deep learning in glioma diagnosis. MRI, magnetic resonance imaging. Datasets Brain Tumor Segmentation (BraTS) Various methods for the segmentation of brain tumors have been presented (23,24). Each of these approaches possesses distinct advantages and contributes meaning-fully to the field. Initially, these methods were employed on images obtained from specific hospitals or radiological laboratories, characterized by their private or public nature, restricting access to other researchers, which limited the potential for method comparison. To address this problem, publicly available datasets and evaluation frameworks were established, allowing for consistent comparisons. Against this backdrop, one dataset of particular interest is the BraTS Challenge dataset, as shown in Table 2, which has the following notable characteristics: Table 2BraTS datasets for glioma segmentation and related tasksYearChallenge/taskTotal images (n)Training images (n)Validation images (n)Test images (n)2014BraTS 2014238200N/A382015BraTS 2015253200N/A532016BraTS 2016391200N/A1912017BraTS 2017477285461462018BraTS 2018542285661912019BraTS 20196263351251662020BraTS 20206603691251662021Adult Glioma2,0401,2512195702022Adult Glioma2,2851,2512198152023Adult Glioma2,0401,2512195702023BraTS-Africa1056015302024Post-Treatment Glioma2,2001,5402204402024BraTS-Africa956035N/AFrom 2014 to 2020, BraTS consisted of a single glioma segmentation track, with each subsequent year progressively expanding upon the same patient cohort and adding new cases; therefore, these rows represent an evolving dataset rather than independent annual collections. The 2021 Adult Glioma dataset served as the foundation for the 2022 and 2023 Adult Glioma tracks; the 2022 test set was expanded from 570 to 815 cases, while the 2023 track retained the original 2021 test set of 570 cases. The 2024 Post-Treatment Glioma task introduced a new post-operative cohort. BraTS, Brain Tumor Segmentation; N/A, not available. BraTS datasets for glioma segmentation and related tasks From 2014 to 2020, BraTS consisted of a single glioma segmentation track, with each subsequent year progressively expanding upon the same patient cohort and adding new cases; therefore, these rows represent an evolving dataset rather than independent annual collections. The 2021 Adult Glioma dataset served as the foundation for the 2022 and 2023 Adult Glioma tracks; the 2022 test set was expanded from 570 to 815 cases, while the 2023 track retained the original 2021 test set of 570 cases. The 2024 Post-Treatment Glioma task introduced a new post-operative cohort. BraTS, Brain Tumor Segmentation; N/A, not available. Image content:The dataset comprises preoperative and postoperative multiparametric MRI scans encompassing T1, T1CE, T2, and T2-FLAIR volumes.Images are sourced from diverse clinical specialties, utilizing both 2D and 3D tools and various scanners with multiple 1.5 T or 3 T mechanisms. Image content: The dataset comprises preoperative and postoperative multiparametric MRI scans encompassing T1, T1CE, T2, and T2-FLAIR volumes. The dataset comprises preoperative and postoperative multiparametric MRI scans encompassing T1, T1CE, T2, and T2-FLAIR volumes. Images are sourced from diverse clinical specialties, utilizing both 2D and 3D tools and various scanners with multiple 1.5 T or 3 T mechanisms. Images are sourced from diverse clinical specialties, utilizing both 2D and 3D tools and various scanners with multiple 1.5 T or 3 T mechanisms. Data preprocessing:The dataset undergoes preprocessing to maintain image integrity, ensuring that the apparent information remains consistent under coordination and standardization conditions.Interpolation is performed at a uniform isotropic resolution of 1 mm³, and skull dissection is executed. Data preprocessing: The dataset undergoes preprocessing to maintain image integrity, ensuring that the apparent information remains consistent under coordination and standardization conditions. The dataset undergoes preprocessing to maintain image integrity, ensuring that the apparent information remains consistent under coordination and standardization conditions. Interpolation is performed at a uniform isotropic resolution of 1 mm³, and skull dissection is executed. Interpolation is performed at a uniform isotropic resolution of 1 mm³, and skull dissection is executed. Dataset evolution:Over time, the BraTS dataset has evolved, initially containing a limited number of clinical images. The challenge of comparing methods based on such a small dataset prompted the need for an expanded image pool.Ground truth images are accurately generated through multiple expert evaluations, mitigating interobserver variability.Four distinct intratumoral structures—edema, enhanced core, non-enhanced (sol-id) core, and necrotic—contribute to creating a realistic background. Dataset evolution: Over time, the BraTS dataset has evolved, initially containing a limited number of clinical images. The challenge of comparing methods based on such a small dataset prompted the need for an expanded image pool. Over time, the BraTS dataset has evolved, initially containing a limited number of clinical images. The challenge of comparing methods based on such a small dataset prompted the need for an expanded image pool. Ground truth images are accurately generated through multiple expert evaluations, mitigating interobserver variability. Ground truth images are accurately generated through multiple expert evaluations, mitigating interobserver variability. Four distinct intratumoral structures—edema, enhanced core, non-enhanced (sol-id) core, and necrotic—contribute to creating a realistic background. Four distinct intratumoral structures—edema, enhanced core, non-enhanced (sol-id) core, and necrotic—contribute to creating a realistic background. Annotation protocol:An annotation protocol is systematically applied, involving manual annotation by expert raters for each case.The segmentation results from all raters are fused to derive a single, consistent segmentation for each case, serving as the ground truth. Annotation protocol: An annotation protocol is systematically applied, involving manual annotation by expert raters for each case. An annotation protocol is systematically applied, involving manual annotation by expert raters for each case. The segmentation results from all raters are fused to derive a single, consistent segmentation for each case, serving as the ground truth. The segmentation results from all raters are fused to derive a single, consistent segmentation for each case, serving as the ground truth. Segmentation criteria:Whole tumor (WT): encompasses all structures within the tumor.Tumor core (TC): includes enhanced tumor, necrotic, and non-enhanced tumor substructures.Enhancing tumor (ET): exclusively covers enhancer structures. Segmentation criteria: Whole tumor (WT): encompasses all structures within the tumor. Whole tumor (WT): encompasses all structures within the tumor. Tumor core (TC): includes enhanced tumor, necrotic, and non-enhanced tumor substructures. Tumor core (TC): includes enhanced tumor, necrotic, and non-enhanced tumor substructures. Enhancing tumor (ET): exclusively covers enhancer structures. Enhancing tumor (ET): exclusively covers enhancer structures. The Cancer Imaging Archive (TCIA) TCIA is a publicly accessible repository that hosts multiple independent imaging collections rather than a single unified dataset. For glioma research, key collections include The Cancer Genome Atlas (TCGA)-GBM and TCGA-LGG (pre-operative MRI with expert segmentation labels), UPENN-GBM (multi-parametric MRI of glioblastoma), UCSF-PDGM (Preoperative Diffuse Glioma MRI with molecular markers), Ivy GAP (Glioblastoma Atlas Project), and LGG-1p19q (lower-grade gliomas with 1p/19q status). These collections differ in imaging protocols, patient populations, and available annotations. This initiative has received financial support from the National Cancer Institute of the United States and is now operated in partnership with the University of Arkansas for Medical Sciences. TCIA serves as a comprehensive resource encompassing more than 14.8 million images and related data, with contributions from numerous cancer research initiatives. Image content:The dataset provides medical image data of various cancer types, including CT, MRI, PET, etc.Imaging data are mainly stored in DICOM format.In addition to image data, TCIA also provides relevant clinical and pathological information, such as patient outcomes, treatment details, genomics and expert analysis. Image content: The dataset provides medical image data of various cancer types, including CT, MRI, PET, etc. The dataset provides medical image data of various cancer types, including CT, MRI, PET, etc. Imaging data are mainly stored in DICOM format. Imaging data are mainly stored in DICOM format. In addition to image data, TCIA also provides relevant clinical and pathological information, such as patient outcomes, treatment details, genomics and expert analysis. In addition to image data, TCIA also provides relevant clinical and pathological information, such as patient outcomes, treatment details, genomics and expert analysis. Characteristics:The dataset provides glioma imaging data in a variety of modalities, mainly including multiparametric magnetic resonance imaging (mpMRI), such as T1-weighted, T2-weighted, T1CE, and T2 FLAIR.Some datasets contain computer-aided and manually corrected tumor segmentation labels. Characteristics: The dataset provides glioma imaging data in a variety of modalities, mainly including multiparametric magnetic resonance imaging (mpMRI), such as T1-weighted, T2-weighted, T1CE, and T2 FLAIR. The dataset provides glioma imaging data in a variety of modalities, mainly including multiparametric magnetic resonance imaging (mpMRI), such as T1-weighted, T2-weighted, T1CE, and T2 FLAIR. Some datasets contain computer-aided and manually corrected tumor segmentation labels. Some datasets contain computer-aided and manually corrected tumor segmentation labels. Annotation protocol:Preliminary segmentation of the image data is performed using a deep learning model to generate a preliminary segmented image of the tumor.The preliminary segmentation results are manually corrected by a professional radiologist or neuroimaging expert to ensure the accuracy and completeness of the annotation.Validate the annotation results to ensure the accuracy and consistency of the annotation.The annotation results are reviewed by several experts to ensure high quality of the annotation. Annotation protocol: Preliminary segmentation of the image data is performed using a deep learning model to generate a preliminary segmented image of the tumor. Preliminary segmentation of the image data is performed using a deep learning model to generate a preliminary segmented image of the tumor. The preliminary segmentation results are manually corrected by a professional radiologist or neuroimaging expert to ensure the accuracy and completeness of the annotation. The preliminary segmentation results are manually corrected by a professional radiologist or neuroimaging expert to ensure the accuracy and completeness of the annotation. Validate the annotation results to ensure the accuracy and consistency of the annotation. Validate the annotation results to ensure the accuracy and consistency of the annotation. The annotation results are reviewed by several experts to ensure high quality of the annotation. The annotation results are reviewed by several experts to ensure high quality of the annotation. Segmentation criteria:Core: usually includes necrotic and enhancing areas of the tumor. In T1CE images, the enhancement area usually shows significant contrast enhancement.Edema: in T2 FLAIR images, edematous areas are usually characterized by high signal intensity.Gross tumor volume (GTV): includes the core and edema regions. Segmentation criteria: Core: usually includes necrotic and enhancing areas of the tumor. In T1CE images, the enhancement area usually shows significant contrast enhancement. Core: usually includes necrotic and enhancing areas of the tumor. In T1CE images, the enhancement area usually shows significant contrast enhancement. Edema: in T2 FLAIR images, edematous areas are usually characterized by high signal intensity. Edema: in T2 FLAIR images, edematous areas are usually characterized by high signal intensity. Gross tumor volume (GTV): includes the core and edema regions. Gross tumor volume (GTV): includes the core and edema regions. TCGA Funded by the National Cancer Institute and the National Human Genome Research Institute, The National Human Genome Research Institute Cancer Research Project provides a large, free reference database for cancer research by collecting and collating various omics data related to cancer. It is important to note that TCGA itself does not host MRI scans; rather, it provides comprehensive molecular characterization including whole-exome sequencing, DNA methylation, and copy number variation (CNV). The imaging data corresponding to TCGA patients is stored in TCIA, and studies using TCGA data typically rely on the TCIA-TCGA fusion linking imaging to genomic annotations. Currently, the repository encompasses data from 33 different cancer types, totaling more than 2 petabytes of information. The data is free and public, which greatly helps cancer researchers to improve the prevention, diagnosis and treatment of cancer. This dataset can also be used for glioma segmentation, grading and genotype classification tasks (25-27). Content:The dataset include genome sequence, transcriptome data (mRNA and miRNA expression), DNA methylation data, CNV data, proteome data, and more.Data levels: TCGA data are categorized into three levels:Level 1: raw sequencing data (e.g. fasta, fastq format).Level 2: aligned bam files.Level 3: processed and normalized data. Content: The dataset include genome sequence, transcriptome data (mRNA and miRNA expression), DNA methylation data, CNV data, proteome data, and more. The dataset include genome sequence, transcriptome data (mRNA and miRNA expression), DNA methylation data, CNV data, proteome data, and more. Data levels: TCGA data are categorized into three levels:Level 1: raw sequencing data (e.g. fasta, fastq format).Level 2: aligned bam files.Level 3: processed and normalized data. Data levels: TCGA data are categorized into three levels: Level 1: raw sequencing data (e.g. fasta, fastq format). Level 1: raw sequencing data (e.g. fasta, fastq format). Level 2: aligned bam files. Level 2: aligned bam files. Level 3: processed and normalized data. Level 3: processed and normalized data. Characteristics:The dataset provides a large amount of genomic data of glioma patients, including DNA sequence, RNA expression, DNA methylation, CNVs, etc.Patient survival time, treatment regimen, and pathologic features are available, and this information is useful for survival analysis and prognostic studies. Characteristics: The dataset provides a large amount of genomic data of glioma patients, including DNA sequence, RNA expression, DNA methylation, CNVs, etc. The dataset provides a large amount of genomic data of glioma patients, including DNA sequence, RNA expression, DNA methylation, CNVs, etc. Patient survival time, treatment regimen, and pathologic features are available, and this information is useful for survival analysis and prognostic studies. Patient survival time, treatment regimen, and pathologic features are available, and this information is useful for survival analysis and prognostic studies. Annotation protocol:The TCGA project has gathered data collection and integration through collaboration with multiple research centers: a large number of cancer samples and related data. These data were integrated and standardized to ensure data consistency and comparability.For each sample, a detailed annotation file (e.g. annotations.txt) is generated to document various attributes and characteristics of the sample.During the annotation process, strict quality control is performed to ensure the accuracy and completeness of the data. Annotation protocol: The TCGA project has gathered data collection and integration through collaboration with multiple research centers: a large number of cancer samples and related data. These data were integrated and standardized to ensure data consistency and comparability. The TCGA project has gathered data collection and integration through collaboration with multiple research centers: a large number of cancer samples and related data. These data were integrated and standardized to ensure data consistency and comparability. For each sample, a detailed annotation file (e.g. annotations.txt) is generated to document various attributes and characteristics of the sample. For each sample, a detailed annotation file (e.g. annotations.txt) is generated to document various attributes and characteristics of the sample. During the annotation process, strict quality control is performed to ensure the accuracy and completeness of the data. During the annotation process, strict quality control is performed to ensure the accuracy and completeness of the data. Synthetic datasets Synthetic datasets are data generated by computer algorithms and simulations, which are widely used in the field of medical imaging, especially in brain tumor segmentation tasks. Its main advantage is that it can generate a large amount of diversified data, which solves the problem of difficult access to real data and privacy protection. Synthetic datasets usually include multimodal MRI data, such as T1, T1CE, T2, and T2 FLAIR, etc., and provide high-quality annotation information covering different subregions of the tumor.IEEE DataPort is an open data-sharing platform to support researchers to upload and share synthetic datasets, e.g., Synthetic Whole-Head MRI BraTS Datasets. There are several points worth noting about this database: Image content and characteristics:Including multimodal MRI images such as T1, T1CE, T2, and T2 FLAIR.Dataset typically contains detailed annotation information covering different subregions of the tumor, such as ET, TC, and WT. These annotated informations provide a benchmark for training and validating segmentation algorithms.Images generated through synthetic techniques can simulate different types of brain tumors and their manifestations in different patients, increasing the diversity and complexity of the dataset and helping to improve the model’s generalization ability.Synthetic MRI technology can provide quantitative measurement values of tissues, such as T1 and T2 relaxation time, proton density, etc., which help to conduct more in-depth tissue characteristic analysis. Image content and characteristics: Including multimodal MRI images such as T1, T1CE, T2, and T2 FLAIR. Including multimodal MRI images such as T1, T1CE, T2, and T2 FLAIR. Dataset typically contains detailed annotation information covering different subregions of the tumor, such as ET, TC, and WT. These annotated informations provide a benchmark for training and validating segmentation algorithms. Dataset typically contains detailed annotation information covering different subregions of the tumor, such as ET, TC, and WT. These annotated informations provide a benchmark for training and validating segmentation algorithms. Images generated through synthetic techniques can simulate different types of brain tumors and their manifestations in different patients, increasing the diversity and complexity of the dataset and helping to improve the model’s generalization ability. Images generated through synthetic techniques can simulate different types of brain tumors and their manifestations in different patients, increasing the diversity and complexity of the dataset and helping to improve the model’s generalization ability. Synthetic MRI technology can provide quantitative measurement values of tissues, such as T1 and T2 relaxation time, proton density, etc., which help to conduct more in-depth tissue characteristic analysis. Synthetic MRI technology can provide quantitative measurement values of tissues, such as T1 and T2 relaxation time, proton density, etc., which help to conduct more in-depth tissue characteristic analysis. Annotation protocol:Before generating synthetic data, it is usually necessary to preprocess real data to extract features and patterns, which will be used to generate synthetic data.Train using deep learning models to generate synthesized MRI data. These synthesized data are similar in appearance and features to real data.After generating synthetic data, annotate to identify different regions of the tumor. The annotation results are usually verified by experts to ensure their accuracy and consistency. Annotation protocol: Before generating synthetic data, it is usually necessary to preprocess real data to extract features and patterns, which will be used to generate synthetic data. Before generating synthetic data, it is usually necessary to preprocess real data to extract features and patterns, which will be used to generate synthetic data. Train using deep learning models to generate synthesized MRI data. These synthesized data are similar in appearance and features to real data. Train using deep learning models to generate synthesized MRI data. These synthesized data are similar in appearance and features to real data. After generating synthetic data, annotate to identify different regions of the tumor. The annotation results are usually verified by experts to ensure their accuracy and consistency. After generating synthetic data, annotate to identify different regions of the tumor. The annotation results are usually verified by experts to ensure their accuracy and consistency. Segmentation criteria:ET: typically manifested as high signal areas in T1CE imaging.TC: includes enhanced and non enhanced TCs, which can be identified in T1CE and T1 weighted imaging.WT: covering enhanced tumor, non enhanced TC, and peritumoral edema (ED), typically manifested as high signal in T2 FLAIR imaging. Segmentation criteria: ET: typically manifested as high signal areas in T1CE imaging. ET: typically manifested as high signal areas in T1CE imaging. TC: includes enhanced and non enhanced TCs, which can be identified in T1CE and T1 weighted imaging. TC: includes enhanced and non enhanced TCs, which can be identified in T1CE and T1 weighted imaging. WT: covering enhanced tumor, non enhanced TC, and peritumoral edema (ED), typically manifested as high signal in T2 FLAIR imaging. WT: covering enhanced tumor, non enhanced TC, and peritumoral edema (ED), typically manifested as high signal in T2 FLAIR imaging. A critical caveat must be acknowledged: synthetic MRI data exhibits a domain gap relative to real clinical scans. Generated images may lack the full complexity of anatomical structures, fail to reproduce scanner-specific noise and artifact patterns, and display distributional shifts in intensity characteristics. Models trained predominantly on synthetic data may suffer from degraded generalization when deployed on real-world clinical data. Performance reported on synthetic test sets should not be directly compared with results obtained on clinical BraTS or TCIA cohorts. Figshare Figshare is one of the many portfolio businesses supported by Digital Science, a division of Macmillian Publishing in the United Kingdom. The objective is to facilitate the discovery of supplementary materials utilized in research. With the support of Digital Science, the Figshare platform has developed into a comprehensive repository for research outputs across nearly 30 disciplines, including biology, medicine, geography, physics, the humanities, and astronomy. The platform’s primary focus is on the biological and medical sciences. Figshare enables researchers to upload a range of materials, including graphs, multimedia, posters, papers (including preprints), multiple files, and data sets. This file sharing mode represents a novel approach to data and information sharing that is not currently available in the context of academic publishing. The utilization of a creative commons license for the dissemination of data serves to mitigate copyright disputes, thereby facilitating global access to and exchange of scientific information. Private datasets Many researchers use private datasets to finish the task of glioma segmentation, grading and genotype classification. Private datasets are usually provided by author’s institution, which are difficult to share due to individual privacy, intellectual property and internal rules (28-38). Comparison and sample images Of the studies surveyed here, the vast majority of investigators used data from BraTS for glioma segmentation task, and some used TCIA, private datasets, and synthetic datasets for their studies. Whereas TCGA does not provide images directly, it provides tumor-related genomic, transcriptomic, and clinical data fused with images to provide additional data support for gene classification tasks. In studies using TCIA and synthetic datasets, most of the images were multimodal synthesized using data from BraTS, including the synthetic whole-head MRI BraTS dataset. This shows that BraTS is particularly important in glioma image processing tasks, and sample images from BraTS and TCIA are provided here for reference, as shown in Figure 2. Sample images. (A) 3D MRI modal T1-weighted (T1), T1-weighted contrast-enhanced (T1CE), T2-weighted (T2), T2 fluid-attenuated inversion recovery (FLAIR) and their corresponding segmentation labels in the BraTS2023 dataset, where black denotes the background, green denotes edema or invagination, red denotes non-enhanced regions, and yellow denotes enhanced regions. (B) Multiparametric magnetic resonance imaging (mpMRI) scans of glioblastoma (GBM) patients from the University of Pennsylvania. Reproduced from (39) under the CC BY 4.0 license. 3D, three-dimensional; MRI, magnetic resonance imaging. Metrics Segmentation task The segmentation task involves the pixel-level classification of entire images, necessitating the categorization of each individual pixel in the image. Each pixel undergoes classification, and the initial dataset posed a challenge due to its limited number of clinical images. Drawing meaningful comparisons between methods based on such a small dataset proved difficult. The feasibility of comparisons improved with the augmentation of the dataset through an increased number of sample images and the meticulous generation of accurate ground truth images. To mitigate interobserver variability, the ground truth was meticulously generated through multiple expert evaluations. The incorporation of four distinct types of intratumoral structures—edema, enhanced core, non-enhanced (solid) core, and necrotic—contributed to the creation of a realistic background. A comprehensive annotation protocol, executed by expert raters, was employed for the manual annotation of each case. Subsequently, the segmentation results from all raters were amalgamated to derive a singular, consistent segmentation regarded as the ground truth for each subject. Currently, within the realm of glioma segmentation, the prevailing quantitative assessment metrics encompass the Dice coefficient, lesion dice, sensitivity, specificity, and Hausdorff coefficient distance. Dice coefficient is a statistical index used to measure the similarity between two sets. It was first used in the fields of biology and ecology, but later it was widely used in text processing, image analysis and other fields, especially in the task of image segmentation. Dice coefficient values range from 0 to 1, and the closer the value is to 1, the higher the similarity between the two sets, that is, the higher the overlap between the predicted results of the model and the real labels, the better the model performance. Lesion Dice is a metric used to evaluate the quality of lesion region segmentation in medical image segmentation. It is an application of the Dice in specific application scenarios (e.g., lesion segmentation) to measure the degree of overlap between predicted and real lesion regions. By using the lesion Dice, researchers and clinicians can more accurately evaluate the performance of medical image segmentation models, thus improving the accuracy of diagnosis and treatment. Sensitivity refers to the proportion of all samples that are actually positive, which are correctly predicted as positive. In the context of image segmentation, sensitivity measures the ability of the model to identify regions of interest (such as diseased regions or other specific anatomical structures). Specifically, sensitivity is defined as the proportion of true positives (TP) to actual positive classes (that is, all realistic interest areas). The higher the sensitivity, it means that the model can capture more actual positive samples, that is, in image segmentation, the model can identify all the regions of interest more completely. This is especially important for medical image segmentation, because in medical diagnosis, missed diagnosis (failure to identify the lesion area) may lead to serious consequences. In practical application, high sensitivity helps to ensure that the model will not miss any important lesion areas, although this may lead to more false positives (FP). Specificity measures the model’s ability to recognize areas of no interest. Specifically, specificity refers to the proportion of samples that are actually negative and are correctly predicted as negative. In the context of image segmentation, this means that the model can accurately identify which regions do not belong to the target area of interest. The higher the specificity, the fewer FPs the model has when identifying non-target areas. High specificity implies a lower FP rate, but it may come at the expense of sensitivity, which may increase the false negative (FN) rate. Therefore, when optimizing the image segmentation model, developers need to find a balance between sensitivity and specificity. The standard calculation method for segmentation task evaluation is the Confusion Matrix, which typically involves the inclusion of three essential components: the object of prediction, true labels, and predicted labels. Consequently, the confusion matrix is stratified into four distinct categories. Based on the actual categories of the true label and the predicted label, the predicted object is categorized into four groups: TP, FP, true negative (TN), and FN. The Dice coefficient, sensitivity, specificity, and Hausdorff distance are formalized as follows: Grading and genotype classification task The commonly employed evaluation metric for grading and genotype classification is the area under the receiver operating characteristic (ROC) curve (AUC). The AUC represents the probability that a randomly chosen positive case receives a higher model score than a randomly chosen negative case, thereby quantifying the classifier’s ability to rank positive instances above negative ones. AUC values range from 0.5 (equivalent to random guessing) to 1.0 (perfect discrimination), with higher values indicating superior classification performance. Another prevalent evaluation criterion is the accuracy rate, denoting “the number of samples predicted correctly divided by the total number of samples”. Accuracy serves as a metric to assess the effectiveness of a classification model by measuring the proportion of correct predictions within the total. In the context of grading and genotype classification, many research works employ AUC and accuracy as primary evaluation indicators. Emerging deep learning paradigms for glioma Vision Transformers (ViT) for glioma analysis The introduction of the ViT marked a paradigm shift in medical image analysis by replacing the inductive biases of CNNs with global self-attention mechanisms. Unlike CNNs, which process images through local receptive fields, ViT divides an image into a sequence of non-overlapping patches and computes pairwise attention across all positions, thereby capturing long-range spatial dependencies that are particularly valuable for glioma segmentation where tumor sub-regions may be spatially distributed and morphologically heterogeneous. Several transformer-based architectures have been specifically developed for brain tumor segmentation. TransBTS was among the first to adapt the transformer for volumetric glioma segmentation, employing a CNN encoder to extract local feature representations followed by a transformer module to model global dependencies across the sequence of encoded tokens. TransUNet integrated transformers into a U-Net-style architecture, using a transformer as the encoder while retaining skip connections to the CNN-based decoder. UNETR extended this approach to full 3D volumes, using a pure transformer encoder that directly processes volumetric patches. Swin UNETR (40) replaced the standard ViT encoder with a hierarchical Swin Transformer that computes self-attention within shifted windows, reducing computational complexity while maintaining effective long-range modeling. As shown in Table 3, Swin UNETR achieved competitive Dice scores on the BraTS 2021 dataset. However, several limitations persist: the computational cost of self-attention scales quadratically with sequence length, and transformers typically require large training datasets. In the medical imaging domain, where annotated datasets are often limited, careful pre-training strategies or hybrid designs remain essential. Table 3Segmentation methods based on deep learningRef.YearDatasetSampleModuleNetworkPerformanceEvaluation-protocol(41)2018BraTS 2018285 scansSDense NetDice: WT 0.79, TC 0.90, ET 0.85Validation(42)2018BraTS 2018285 scansSESPNetDice: WT 0.850, TC 0.782, ET 0.665Validation(43)2018BraTS 2018285 scansSU-NetDice: WT 0.871, TC 0.771, ET 0.706Validation(44)2018BraTS2018285 scansSResNet†Dice: WT 0.941, TC 0.914, ET 0.835Validation(45)2018BraTS 2018285 scansMCascaded U-NetDice: WT 0.8980, TC 0.8118, ET 0.7519Validation(46)2018BraTS 2018285 scansMCascaded V-NetDice: WT 0.9048, TC 0.8364, ET 0.7748Validation(47)2018BraTS 2018285 scansMCascaded U-NetDice: WT 0.878, TC 0.785, ET 0.720Validation(48)2018BraTS 2018285 scansM2D-3D ensemble CNNDice: WT 0.83, TC 0.72, ET 0.69Validation(49)2018BraTS 2018285 scansSU-NetDice: WT 0.85, TC 0.72, ET 0.66Validation(50)2018BraTS 2018351 scansSCNNDice: WT 0.868, TC 0.805, ET 0.783Validation(51)2019BraTS 2019335 scansMCascaded U-NetDice: WT 0.8806, TC 0.8084, ET 0.8029Validation(52)2019BraTS 2019335 scansMCascaded U-NetDice: WT 0.88796, TC 0.83697, ET 0.83267Validation(53)2019BraTS 2019335 scansMCascaded MECU-NetDice: WT 0.902, TC 0.824, ET 0.777Validation(54)2019BraTS 2019335 scansMCascaded U-Net†Dice: WT 0.90712, TC 0.83911, ET 0.77338Validation(55)2019BraTS 2019335 scansM2D-3D ensemble CNNDice: WT 0.8865, TC 0.7372, ET 0.7743Validation(56)2019BraTS 2019335 scansM2D-3D ensemble CNNDice: WT 0.89, TC 0.83, ET 0.81Validation(57)2019BraTS 2019335 scansSGANDice: WT 0.8965, TC 0.7901, ET 0.7665Validation(58)2019BraTS 2019335 scansSU-NetDice: WT 0.737, TC 0.862, ET 0.751Validation(59)2020BraTS 2015, BraTS 2017569 scansSResNet†15 Dice: WT 0.9401, TC 0.9429, ET 0.8732; 17 Dice: WT 0.9463, TC 0.856, ET 0.9354Internal test(60)2020BraTS 2013, BraTS 2015, BraTS 2018624 scansSAFPNetDice: WT 0.8658, TC 0.7688, ET 0.74434Internal test(61)2020BraTS 2018, Private582 scansSU-netDice: WT 0.92, TC 0.90, ET 0.81External test(62)2020BraTS 2017, Private315 scansSFCNDice: WT 0.8899, TC 0.8134, ET 0.7394External test(63)2020BraTS 2017285 scansSV-NetDice: WT 0.89, TC 0.80, ET 0.74Validation(64)2021BraTS 2018, Private285 scansSERV-NetDice: WT 0.818, TC 0.9121, ET 0.8662External test(65)2021BraTs 2019349 scansSMLK CNNDice: WT 0.90, TC 0.83, ET 0.74Validation(66)2021BraTS 2017, BraTS 2018682 scansSDDU-NetDice: WT 0.8986, TC 0.8388, ET 0.7837Validation(67)2021BraTS 20211,251 scansSCMM-NetDice: WT0.8351, TC 0.8060, ET 0.7557Validation(68)2021BraTS 20211,251 scansMResNet, U-NetDice: WT 0.8921, TC 0.8130, ET 0.8064Validation(69)2021BraTS 20211,251 scansSnnU-Net†Dice: WT 0.9300, TC 0.8843, ET 0.8760Validation(70)2021BraTS 20211,251 scansSAttention U-NetDice: WT 0.913, TC 0.869, ET 0.858Validation(71)2021BraTS 20211,251 casesMCascaded U-NetDice: WT 0.8884, TC 0.8197, ET 0.7502Validation(72)2021BraTS 20211,251 casesMEnsemble CoTrHRNetDice: WT 0.9256, TC 0.8774, ET 0.8576Validation(73)2021BraTS 20211,251 casesMEnsemble DeepSeg nnU-NetDice: WT 0.8410, TC 0.8733, ET 0.9200Validation(74)2021BraTS 2019335 scansSFCNDice: WT 0.89, TC 0.78, ET 0.76Validation(75)2021BraTS 2018285 scansSU-NetDice: WT 0.829, TC 0.749, ET 0.591Validation(76)2021BraTS 2018, BraTS 2019620 scansSDFP-ResU-Net18 Dice: WT 0.8970, TC 0.9068, ET 0.8431; 19 Dice: WT 0.9028, TC 0.8453, ET 0.7985Validation(77)2021BraTS2015274 scansSU-NetDice: WT 0.82, TC 0.68, ET 0.61Internal test(78)2021BraTS 20211,251 casesSU-NetDice: WT 0.922, TC 0.868, ET 0.830Validation(79)2021BraTS 20212,140 casesSU-NetDice: WT 0.73, TC 0.67, ET 0.63Internal test(80)2021BraTS 20211,251 casesSSegNetDice: WT 0.928, TC 0.878, ET 0.847Validation(77)2021BraTS 2015200 scansSTIU-NetsDice: WT 0.85, TC 0.69, ET 0.63Validation(81)2021BraTS 2018285 scansMCascaded CNNDice: WT 0.92, TC 0.91, ET 0.87Validation(82)2021BraTS 20211,251 casesSU-NetDice: WT 0.81389, TC 0.77, ET 0.8539Validation(83)2021BraTS 20211,251 casesSU-NetDice: WT 0.8907, TC 0.8073, ET 0.7802Validation(40)2021BraTS 20211,251 casesSSwin UNETRDice: WT 0.926, TC 0.885, ET 0.858Validation(84)2022BraTS 2019335 scansSMIRA U-NetDice: WT 0.885, TC 0.879, ET 0.818Validation(85)2022BraTS 2018285 patientsSAttention FCNDice: WT 0.871, TC 0.789, ET 0.741Validation(86)2022BraTS 2019335 scansSAttention ResUNetDice: WT 0.911, TC 0.876, ET 0.801Validation(87)2022BraTS 2017285 scansSAttention ResNetDice: WT 0.743, TC 0.795, ET 0.882Validation(88)2022BraTS 2019335 scansSU-NetDice: WT 0.91, TC 0.72, ET 0.82Validation(89)2022BraTS 2019335 scansSU-NetDice: WT 0.9215, TC 0.8671, ET 0.8349Validation(90)2022BraTS 2019335 scansSU-Net†Dice: WT 0.959, TC 0.943, ET 0.892Validation(91)2022BraTS 2019335 scansMCascaded 3D U-Net, 3D U-Net++Dice: WT 0.890, TC 0.842, ET 0.835Validation(91)2022BraTS 2018285 scansMEnsemble CNN and ASCNNDice: WT 0.892, TC 0.763, ET 0.772Validation(92)2022BraTS 2018, BraTS 2019, BraTS 20201,038 scansMCascaded U-NetDice: WT 0.9045, TC 0.843, ET 0.8216Internal test(93)2022BraTS 2017, BraTS 2019, BraTS 2020989 scansMCascaded MRAU-NetDice: WT 0.9018, TC 0.8722, ET 0.8674Validation(94)2022TCIA439 scansSCNN-TransformerDice: WT 0.9071, TC 0.786, ET 0.7589External test(95)2023BraTS 2018, BraTS 2019620 scansSCB-D²RNet†18 Dice: WT 0.982, TC 0.987, ET 0.976; 19 Dice: WT 0.983, TC 0.989, ET 0.977Validation(96)2023BraTS 2019, BraTS 2020831 scansSU-Net19 Dice: WT 0.7803, TC 0.8831, ET 0.8172; 20 Dice: WT 0.7675, TC 0.8925, ET 0.8110Internal test(97)2023BraTS 20211,251 casesSVITDice: WT 0.826, TC 0.799, ET 0.876Validation(98)2023BraTS 2019335 scansSU-NetDice: WT 0.8149, TC 0.9097, ET 0.8581Validation(99)2023BraTS 2018, BraTS 2019620 scansSAttention U-NetDice: WT 0.8476, TC 0.8641, ET 0.7810Validation(100)2024BraTS 20231431 casesMnnU-NetBraTS-GLI Dice: WT 0.90, TC 0.81, ET 0.79; BraTS-PED Dice: WT 0.85, TC 0.69, ET 0.60Challenge test(101)2024BraTS2020750 casesMGFNetDice: WT 0.8579, TC 0.8113, ET 0.8064Internal test(102)2024BraTS2019, BraTS2020954 casesMKFD-Net19 Dice: WT 0.756, TC 0.900, ET 0.805; 20 Dice: WT 0.760, TC 0.899, ET 0.804Validation(103)2024BraTS2018-20211,589 casesMResMHA-Net†Dice: WT 0.93, TC 0.92, ET 0.91Internal test(104)2025BraTS 2021 TCGA-TCIA110 casesSEfficient U-NetDice: WT 0.9132, TC 0.9254, ET N/AExternal test(105)2025BraTS 2020418 casesMGSG-UNetDice: WT 0.9128, TC 0.8582, ET 0.8588Internal test(106)2025BraTS 2020494 patientsSBTS U-NetDice: WT 0.908, TC 0.878, ET 0.811Validation(107)2025BraTS 2023285 casesS3D U-NetDice: WT 0.9168, TC 0.8954, ET 0.8674Internal test†, the network models that perform better each year for reference. AFPNet, a 3D fully convolutional neural network with atrous-convolution feature pyramid; Attention FCN, attention fully convolutional network; Attention ResNet, attention residual network; Attention ResU-Net, attention residual U-net; BraTS, Brain Tumor Segmentation; BTS U-Net, brain tumor segmentation U-net; Cascaded MECU-Net, cascaded memory-efficient cascade 3D U-net; Cascaded MRAU-Net, cascaded multi-scale residual attention U-net; CB-D²RNet, context bridge-dense dilated residual net; CMM-Net, contextual multi-scale multi-level network; CNN, convolutional neural network; 2D, two-dimensional; 3D, three-dimensional; DDU-net, dual dense U-structure network; Dense Net, densely connected convolutional network; DFP-ResU-Net, dilated convolutional feature pyramid residual U-net; CoTrHRNet, contextual transformer HRnet; DeepSeg nnU-Net, deepseg no-new-net; ERV-Net, efficient 3D residual neural network; ESPNet, efficient spatial pyramid of dilated convolutions; ET, enhancing tumor; FCN, fully convolutional network; GAN, generative adversarial network; GFNet, global feature network; GSG-UNet, glioma segmentation and grading U-net; KFD-Net, knowledge fusion decision network; M, multi-module; MIRA U-Net, multi-modal instruction-guided restoration architecture U-net; nnU-Net, no-new-net; Ref., reference; ResMHA-Net, resnet multi-head attention u-net; ResNet, residual neural network; S, single-module; SegNet, segmentation network; Swin UNETR, swin unet transformers; TC, tumor core; TCGA, The Cancer Genome Atlas; TCIA, The Cancer Imaging Archive; TIU-Nets, triple intersecting U-nets; VIT, vision transformer; WT, whole tumor. Segmentation methods based on deep learning †, the network models that perform better each year for reference. AFPNet, a 3D fully convolutional neural network with atrous-convolution feature pyramid; Attention FCN, attention fully convolutional network; Attention ResNet, attention residual network; Attention ResU-Net, attention residual U-net; BraTS, Brain Tumor Segmentation; BTS U-Net, brain tumor segmentation U-net; Cascaded MECU-Net, cascaded memory-efficient cascade 3D U-net; Cascaded MRAU-Net, cascaded multi-scale residual attention U-net; CB-D²RNet, context bridge-dense dilated residual net; CMM-Net, contextual multi-scale multi-level network; CNN, convolutional neural network; 2D, two-dimensional; 3D, three-dimensional; DDU-net, dual dense U-structure network; Dense Net, densely connected convolutional network; DFP-ResU-Net, dilated convolutional feature pyramid residual U-net; CoTrHRNet, contextual transformer HRnet; DeepSeg nnU-Net, deepseg no-new-net; ERV-Net, efficient 3D residual neural network; ESPNet, efficient spatial pyramid of dilated convolutions; ET, enhancing tumor; FCN, fully convolutional network; GAN, generative adversarial network; GFNet, global feature network; GSG-UNet, glioma segmentation and grading U-net; KFD-Net, knowledge fusion decision network; M, multi-module; MIRA U-Net, multi-modal instruction-guided restoration architecture U-net; nnU-Net, no-new-net; Ref., reference; ResMHA-Net, resnet multi-head attention u-net; ResNet, residual neural network; S, single-module; SegNet, segmentation network; Swin UNETR, swin unet transformers; TC, tumor core; TCGA, The Cancer Genome Atlas; TCIA, The Cancer Imaging Archive; TIU-Nets, triple intersecting U-nets; VIT, vision transformer; WT, whole tumor. Foundation models for glioma segmentation The SAM (108), released by Meta AI in 2023, represents a watershed moment in computer vision by introducing a promptable foundation model for image segmentation. SAM comprises three core components: a Vision Transformer-based image encoder, a lightweight prompt encoder, and a mask decoder. The application of SAM to glioma segmentation, however, reveals significant challenges stemming from the domain gap between natural images and medical MRI scans. In zero-shot evaluation on brain tumor MRI, SAM frequently produces inaccurate segmentations because its training data largely excluded medical images. To bridge this domain gap, MedSAM was developed as the first large-scale medical segmentation foundation model, fine-tuned on over one million medical image-mask pairs. For glioma analysis, several adaptation strategies have been explored, including full fine-tuning and parameter-efficient alternatives such as low-rank adaptation (LoRA). Beyond direct segmentation, SAM and its medical variants have emerged as valuable annotation assistance tools, where radiologists provide point or bounding box prompts to refine model-generated masks, significantly reducing the manual annotation burden. Diffusion models and SSL Diffusion models have emerged as a powerful class of generative models with significant potential for medical image synthesis and data augmentation. Unlike GANs, which train a generator and discriminator in an adversarial minimax game, diffusion models learn to reverse a gradual noising process. Denoising diffusion probabilistic models (DDPMs) and their latent-space variants have demonstrated the ability to generate high-fidelity synthetic MRI scans with realistic tumor appearances. For glioma analysis, diffusion models have been applied to generate synthetic multi-modal MRI scans, thereby augmenting limited training datasets. SSL offers an alternative pathway to reduce reliance on manually annotated data by pre-training models on large corpora of unlabeled medical images. Contrastive learning methods such as SimCLR and MoCo learn discriminative representations by maximizing agreement between differently augmented views of the same image. Masked autoencoders (MAE) represent another influential SSL paradigm, randomly masking input image patches and training the model to reconstruct the missing content. The combination of SSL pre-training on large unlabeled medical datasets with fine-tuning on limited annotated glioma data represents a promising strategy to improve model performance while reducing the burden of expert annotation. Segmentation Accurate and fast determination of glioma boundary is critical (6). Deep learning-based glioma segmentation methods provide a solution to the issues associated with manual data analysis and the selection of segmentation features in traditional machine learning algorithms. The issue of data imbalance represents a significant challenge in the context of glioma segmentation, with the potential to significantly influence the accuracy of the resulting segmentation. The primary factors contributing to this issue are imbalances in pixels, patients, and labeling biases. In particular, discrepancies in pixel distribution across distinct sub-regions of glioma, an overrepresentation of HGG relative to LGG, and disparate standards among different experts interpreting the same MRI scans contribute to data imbalances. The application of automated glioma segmentation through deep learning-based methods effectively addresses the limitations associated with strong prior constraints and manual intervention in the segmentation process, thereby enhancing the robustness of the approach. This results in superior segmentation performance, particularly in the context of large-scale, multi-modal, and complex glioma segmentation scenarios (109). Initially, many approaches for brain tumor segmentation concentrate on refining neural network to facilitate stable training and the acquisition of informative, discriminative features, ultimately achieving precise segmentation. Several studies (43,110-112) utilized single-path networks or single model for the consideration of computational efficiency. A voxel-wise class probability prediction (113) used separate 3D CNNs to classify HGG and LGG images. Lopez et al. (4) introduced a dilated residual network for patch-based training, where equiprobable patches are fed into the network training. McKinley et al. (114) proposed an encoder-decoder FCN based architecture to segment diverse tumor sub-regions. As the depth of CNN increases, accommodating more layers in the network poses challenges such as gradient explosion or gradient disappearance during the training process. To address this issue, Rivera et al. (115) integrated residual connections from brain segmentation networks into the module design. Residual connections mitigate vanishing and exploding gradients by incorporating the input of a convolutional module into its output, thus avoiding degradation, expediting convergence, and enhancing the accuracy. Presently, residual connectivity has evolved into a standard for designing modules and operating architectures in deep networks. DenseNet and residual connections were also adopted for deep networks, which looks more conducive to gradient backpropagation (116,117). A noteworthy variant of the FCN is the U-Net architecture, serving as an archetypal segmentation backbone network. Researchers strive to augment the efficacy of neural networks by incorporating advantageous techniques into the backbone network. These techniques encompass the expansion of the receptive field, integration of attention mechanisms, and fusion of distinct features to enhance overall network performance. Dong et al. (118) modified U-Net by employing 3D input, in which the depth of the network was reduced to 3. Isensee et al. (9) reported a modified U-Net employing dice loss functions and extensive data augmentation to avoid over-fitting. Compared with single-path networks, multi-path networks can extract different features from different paths at different scales. The extracted features are combined (added or concatenated) together for further processing. A common explanation is that a large-scale path, or a path with a large-sized kernel or input, allows the network to learn global features. While small-scale paths (paths with small-sized kernels or inputs) drive the network to learn local features. Global features tend to provide global information such as tumor location, size, and shape, while local features provide descriptive details such as tumor texture and boundaries. Havaei et al. (5) proposed a voxel-based approach, in which a multipath CNN architecture was used for segmentation. One path uses a 32×32 sized 2D patch and the other uses a 5x5 sized fully connected input with the same central pixel. Naceur et al. (119) extended the network structure by using the overlapping patch prediction method, in which the center of the target patch was associated with the adjacent overlapping patches. In recent years, researchers have increasingly adopted ensemble approaches to address these challenges. Ensemble methodologies can be classified into two principal categories: serial and parallel approaches. In the serial approach, multiple networks are structured in a cascade, comprising upstream and downstream networks. The output of an upstream network serves as the input for the downstream network, thereby establishing a sequential dependency. In contrast, the parallel approach involves the concurrent operation of multiple networks, which may receive either identical or disparate inputs, thereby enabling the extraction of comprehensive details collectively. The final output is determined through a process of majority voting or averaging of all network outputs. Generally, cascaded networks have been proposed to mitigate data imbalance challenges. For instance, a 2D Input Cascaded CNN and a cascaded FCN with residual connections were introduced to sequentially segment WTs and internal tumor regions (53). Su et al. (26) also combined multiple networks, such as WNet, TNet, and ENet, in a cascaded fashion, where individual network outputs contributed to the final segmentation through an averaging process. Xu et al. (120) proposed a cascaded U-Net with three networks. The DeepMedic, the FCN, and the U-net as well as mutations of these three networks were integrated into an ensemble of seven networks (121). Jiang et al. (52) proposed an end-to-end two-stage cascaded U-net. In the first stage, a variant of U-net is used to train a coarse prediction. In the second stage, the network’s width is augmented, and two decoders are utilized to enhance overall performance. However, these strategies all depend on individual network performance, and the increased computing resource consumption is proportional to the number of network candidates. Diverse MRI modalities emphasize different characteristics of gliomas. Multi-modal imaging plays a key role in medical image analysis and application. Emphasis on different organizations. Effective use of multimodal information is one of the key factors for brain tumor segmentation based on MRI. Based on the completeness of the available modalities, we divided multimodal brain tumor segmentation into two scenarios: multimodal based information utilization and limited information processing with missing modalities. The effective integration of multi-modality information (76) constitutes a significant advancement in tumor segmentation. Notably, Li et al. (122) investigated the inter-modality relationships by pairing each two modalities and directing them to the downstream network. Havaeia et al. (5) integrates four networks to extract features from individual modalities and then concatenates the discriminated modality-aware features. Zhang et al. (109) pursued the fusion of features through complex cross-modal and single-modal feature pairing modules. Additionally, some networks utilize attention mechanisms to extract multimodal information. The fusion module combines spatial and channel attention mechanisms in order to identify and emphasise the most valuable features, while simultaneously suppressing any redundant ones. More recently, Lv et al. (123) proposed MamTrans, which integrates attention mechanisms with state-space models for the MRI segmentation of HGGs and brain meningiomas, illustrating the ongoing evolution of segmentation architectures beyond conventional convolutional designs. Incorporating a customized loss function to modulate gradients during the training of a segmentation model proves to be a viable strategy for enhancing glioma segmentation performance. A customized cross-entropy loss specifically for boundary pixels has been proposed in combination with an auxiliary task to solve the boundary localization problem (6,41). Liu et al. (124) proposed a regularization method using the reconstruction task to apply the loss function to improve the accuracy of pixel reconstruction. In addition, the spatial loss function can be utilized to maximize the retention of spatial information in the learned features. For the imbalance problem, multi-class dice loss can be used, i.e., the error weight in the backpropagation process is inversely proportional to the proportion of the classes, which effectively solves the problem of unbalanced datasets (9,125). Customized loss functions serve to attenuate the influence of easy-to-classify samples in the loss function, thereby mitigating the impact of gradient bias during learning from imbalanced datasets (126). Table 3 presents a selection of glioma segmentation methods that have been developed in recent years based on multiple networks. Among the deep learning based glioma segmentation methods surveyed in this paper, several network models have demonstrated strong performance, including ResNet, U-Net and its variants, CB-D²RNet and ResMHA-Net. However, direct comparison across different BraTS editions and evaluation protocols should be interpreted with caution, as differences in dataset composition, preprocessing, and validation schemes can substantially affect reported performance. CB-D²RNet reported high Dice scores on BraTS 2018/2019, but these results may not be directly comparable to methods evaluated on different challenge editions. The choice of model should be guided by the specific application context, available data, and computational constraints rather than leaderboard ranking alone. Grading Radiological features constitute a crucial determinant for the grading of gliomas, serving as a pivotal tool in assessing the severity of the condition. For glioma images, a set of shareable feature vectors can be extracted and encoded in the glioma imaging phenotype using deep learning models (127). Subsequently, a classifier based on logistic regression demonstrated a 90% accuracy in predicting glioma grades. Wiem et al. (128) drew inspiration from wound healing processes and proposed an efficient Network In-Network (NIN), a neural network distinguished by systematic radiomics image features for class prediction. This model achieved an accuracy of 93% across a cohort of 285 glioma patients. Gutta et al. (129) employed a customized CNN to predict glioma grades, comparing the proposed model with support vector machines, random forests, and gradient boosting, all trained with radiomics features. The CNN-learned features outperformed methods solely reliant on radiomics features. The integration of deep learning with radiomics using contrast-enhanced T1-weighted imaging represents a promising approach for glioma grading (31). The integration of deep learning with radiomics features, in conjunction with optimal classification models, has demonstrated substantial efficacy in accurately grading gliomas and facilitating informed clinical decision-making. An apparent advantage distinguishing CNN from traditional machine learning lies in the CNN’s intrinsic ability to autonomously learn discriminative features, particularly valuable in the context of glioma grading. In previous studies, Al-Zurfi et al. (130) developed a CNN based DNN, integrating an extensive iteration matrix involving eleven distinct machine learning algorithms. The model demonstrated remarkable sensitivity, specificity, and accuracy, reaching 100%, 90%, and 93.3%, respectively. Tandel et al. (131) leveraged five pre-trained CNNs, AlexNet, VGG16, ResNet18, GoogleNet, and ResNet50, into a classifier by employing a voting mechanism to differentiate between HGG and LGG. This adaptable model is designed to select the optimal CNN for classification, highlighting the CNN’s capacity to extract pertinent features across diverse image datasets. Decuyper et al. (30) devised an automated MRI-based pipeline for segmentation and grade prediction, employing a 3D U-Net for segmentation and feature extraction, coupled with a CNN-based classifier for glioma grading. This system exhibited commendable performance on the TCIA dataset, achieving an area under the curve (AUC) of 93%. Sultan et al. (132) implemented a CNN model with rectified linear unit (ReLU) as the activation function and data augmentation techniques for brain tumor grading, achieving an impressive accuracy of 96.13% and 98.7%. Zhuge et al. (133) introduced a 2D mask R-CNN model and a 3DConvNet model for glioma grading, evaluated on the TCIA and BraTS 2018 datasets. Results indicated superior accuracy with the 3D model, albeit at a slightly longer processing time. Sameer et al. (3) proposed an automated system utilizing a 3D-CNN for brain tumor classification into HGG and LGG, achieving an impressive Dice similarity coefficient (DSC) of 99.7% on the BraTS 2015 dataset. These studies collectively underscore the efficacy of CNN-based models in automating and enhancing glioma grading tasks. In addition to CNN, several typical DNNs also exhibit notable efficacy in glioma grading. Long short-term memory (LSTM) networks, renowned for their proficiency in mitigating the vanishing gradient problem through memory preservation, have demonstrated competitive performance in this domain. For instance, Shahzadi et al. (134) employed a VGG cascade integrated with an LSTM network to discern between HGG and LGG. The outcomes indicated superior classification accuracy achieved by the combination. Meanwhile, the integration of diverse models through a fusion scheme serves to make use of their respective advantages, thereby improving overall performance. The utilization of multi-modal imaging offers the potential to enhance model performance by providing complementary information. However, the concurrent design of an appropriate network and effective utilization of multi-modal data pose significant challenges. Ge et al. (135) have integrated multi-scale feature learning into a deep convolutional network architecture, introducing a novel feature fusion scheme to discern between HGG and LGG. Ye et al. (136) proposed an end-to-end 3D CNN with a gated multimodal unit (GMU) fusion, allowing integration of information across three dimensions and multiple modalities. The proposed 3D CNN with GMU fusion exhibited notable performance in distinguishing benign gliomas from malignant gliomas. Chen et al. (137) introduced a cross-fusion network (CFNet) designed for high- and LGG classification without the need for segmentation operations, thereby significantly saving time and enhancing efficiency. Transfer learning often leverages information acquired from a previously trained model to address problems of a similar nature. Transfer learning models exhibit accelerated learning, reduced time consumption, and improved results while demanding less training data. Naser (138) integrated a fully connected classifier with transfer learning based on a pre-trained VGG16 convolutional base for glioma grading. The grading model demonstrated accuracy, sensitivity, and specificity of 0.95, 0.97, and 0.98, respectively, in a cohort of 110 patients with LGG classified into grade II and grade III. Tripathi (139) introduced a novel Dempster-Shafer Theory (DST), employing four residual networks to enhance grading performance through transfer learning. Experimental results showed the effectiveness of DST in LGG grading. Yang (28) conducted training experiments with GoogleNet and AlexNet using transfer learning, and training from scratch on a dataset of 113 glioma patients. Both AlexNet and GoogLeNet exhibited improved performance with transfer learning and fine-tuning. Ahuja et al. (140) employed a VGG19 transfer learning model to classify gliomas into HGG, LGG and normal cases, achieving an AUC of 0.99. This demonstrates the effectiveness and potential of transfer learning in medical image analysis and glioma grading, especially in clinical settings dealing with small datasets and the need for fast and accurate diagnosis. The generative adversarial network (GAN) technique has been introduced to address the issue of insufficient data in brain MRI, facilitating the real-time generation and synthesis of data across diverse modalities. Han et al. (141) introduced the progressive growing GAN (PGGAN) model, incorporating features extracted from DNN into an unsupervised image-image translation process. This translation aims to generate realistic images for tumor detection and grading. Kermi et al. (142) applied a deep learning-based three-dimensional generative adversarial network (3D-GAN) to automatically discriminate and segment sub-regions of HGG and LGG. Song (33) proposed a glioma MRI super-resolution reconstruction method based on a channel attention generative adversarial network (CGAN), enhancing the clarity of MRI images and enabling experts to accurately detect and grade gliomas. Wang et al. (99) addressed the constraint of insufficient datasets by employing a combination of variational autoencoders (VAEs) and GAN within an unsupervised deep generative framework. The ResNet50 classifier achieved performance metrics with values of 0.769, 0.837, 0.833, and 0.80 for recall, specificity, precision, and F1 score, respectively. Kobayashi et al. (107) utilized a CNN model for feature extraction, incorporating GAN for data augmentation, resulting in an accuracy of 91.7%. The attention mechanism can be conceptualized as a dynamic selection process, wherein features are adaptively weighted based on their input importance. This mechanism assumes an increasingly significant role in glioma grading. Jeong et al. (143) identified that an attention-based fusion network, coupled with adversarial learning, serves as a viable solution for addressing the issue of missing modality features. The approach achieved an AUC of 87.76% in glioma grading. Wu et al. (144) proposed a novel MRI-oriented attention-based glioma grading network (AGGN) integrating both channel and spatial attention mechanisms to extract shallow and deep features from each MRI modality, demonstrating high generalization ability and robustness. Tripathi et al. (145) incorporated two novel CNN architectures into their deep learning model. The first model, based on 3D U-net with spatial and channel attention mechanisms, was designed for glioma segmentation. The second model adopted a supervised multitask learning approach based on CNN for classification and grading. Extensive experimental results proved the reliability of the proposed framework. Some new detection technologies also play a positive role in glioma grading. The microvascularity of gliomas holds a strong correlation with their grading. Li (29) recognized this association and proposed an automated deep learning-based method to extract features from complex microvascular networks in large hematoxylin and eosin (H&E) stained specimens. The results indicated that quantified microvascular features extracted by the deep learning model hold potential for glioma grading. In addition, near-infrared fluorescence imaging offers opportunities for rapid and accurate real-time diagnosis. Xiao et al. (32) introduced a novel neural architecture search method with two learnable stems for multi-modal low-level feature fusion. In a dataset comprising 1,115 surgical glioma specimens, this method achieved an AUC of 0.843 and an accuracy of 0.634. Moreover, diffusion tensor imaging (DTI) captures pathophysiological processes induced by glioma. Bacchi et al. (146) employed a CNN model to classify HGG using DWI, FLAIR, and post-contrast T1 sequences. The model combining DWI + FLAIR sequences achieved the highest accuracy of 0.82. Vidyadharan et al. (147) employed DTI to classify LGG from HGG with high accuracy. A brain white matter (WM) network connectivity matrix was designed to extract features, and a deep CNN was employed for classification. Beyond grading, deep learning has also been applied to closely related differential diagnostic tasks. For example, Xiong et al. (148) developed deep learning models for the rapid discrimination of HGGs from solitary brain metastases using multi-plane T1CE images, a distinction that frequently poses a challenge in routine clinical practice. Due to the importance of datasets for deep learning-based glioma grading, Figure 3 illustrates the proportion of datasets used in the glioma grading studies surveyed in this paper. Of these, the BraTS database occupies an absolute dominant position of 52% and is widely recognized for its high quality and detailed annotation. Private databases accounted for 23%, showing researchers’ emphasis on specific needs and data diversity; TCIA and TCGA were used at 12% and 7%, respectively, and the multimodal data they provided facilitated in-depth cross-disciplinary research. Kaggle and Figshare, although accounting for a smaller share of the total, were still used as complementary sources to drive innovation and rapid validation. Summary of the proportion of databases used for glioma grading. TCGA, The Cancer Genome Atlas; TCIA, The Cancer Imaging Archive. Overall, the diversity of these databases provides a rich material and research base for deep learning studies on glioma grading. Table 4 provides some typical grading methods based on deep learning. Table 4Grading methods based on deep learningRef.YearDatasetSampleModuleNetworkPerformanceEvaluation-protocol(136)2017BraTS 2015270 patientsSGMUAccuracy =82.1%Validation(135)2018BraTS 2017285 patientsSCNNAccuracy =89.47%Validation(130)2018TCIA30 patientsMEnsemble CNN†Accuracy =93.3%External test(134)2018BraTS 201560 patientsSLSTM VGG-16Accuracy =84%Internal test(28)2018Private113 patientsMAlexNet, GoogLeNetAUC =0.939External test(132)2019TCIA73 casesSCNN†Accuracy =98.7%External test(141)2019BraTS 2016220 patientsMGAN CNNAccuracy =96.70%Validation(29)2019Private350 patientsSFCNAUC =0.953External test(140)2020BraTS 201942 patientsMVGG-19 Transfer learningAUC =0.940Internal test(134)2020BraTS 2018315 patientsSU-net†Accuracy =97.1%Validation(129)2021BraTS 2017237 patientsSCNNAccuracy =87.0%Validation(127)2021BraTS 2019335 patientsSLogistic regression CNNAUC =0.951Validation(30)2021BraTS 2019, Private738 patientsSU-NetAccuracy =94.0%External test(131)2022TCIA112 patientsSCNN†Accuracy =99.06%External test(137)2022BraTS 2020194 casesMCross Fusion CNNAUC =0.9769Internal test(143)2022BraTS 2017285 patientsSAttention GANAUC =0.945Validation(149)2022BraTS 2018, Figshare263 patientsMCapsuleNet, GANAUC =0.949External test(142)2022BraTS 20221,470 casesSGANAUC =0.955Validation(139)2022TCIA163 patientsMResNetAUC =0.941External test(31)2022Private, TCIA151 patientsSVGG-16AUC =0.950External test(128)2022BraTS 2017285 patientsSNINAccuracy =93.0%Validation(32)2022Private1,115 slicesSDLS-DARTSAUC =0.954External test(147)2022Private92 patientsSGRAD CNN†AUC =0.956External test(150)2022Private1,426 slicesSResNetAUC =0.942External test(144)2023BraTS 2018, BraTS 2019326 patientsSAttention CNNAUC =0.947Internal test(145)2023BraTS 2019335 patientsSAttention CNNAUC =0.946Validation(26)2023TCGA507 patientsSResNetAUC =0.943External test(25)2023TCGA402 casesSSASG-GCNAccuracy =93.62%External test(27)2023TCGA737 patientsSTDC-DistillAUC =0.9343External test(33)2023Private230 patientsSAttention ResNetAUC =0.944External test(151)2024BraTS 2018285 casesMU-Net and VGG-16Accuracy =97.89%Validation(152)2024BraTS 2019365 casesSResNet101Accuracy =95.52%Validation(153)2024BraTS 2018-2020989 casesMResNet50Accuracy =98.92%Validation(154)2024BraTS 2019369 casesMResMT†AUC =0.9953Validation(155)2025BraTS2019335 patientsMEfficientNetB2Accuracy =98.49%Validation(156)2025RIDER, REMBRANDT, TCGA-LGG4,570 casesMResNet-50Accuracy =97.89%External test(157)2025BraTS 2018, BraTS 2020654 patientsMHybrid integrated network†Accuracy =99.35%Validation†, the network models that perform better each year for reference. The network models that perform better each year are bolded in the table for reference. Attention ResNet, attention residual network; AUC, area under the curve; BraTS, Brain Tumor Segmentation; CapsuleNet, capsule network; CNN, convolutional neural network; DLS-DARTS, double-learnable-stem differentiable architecture search; FCN, fully convolutional network; GAN, generative adversarial network; GMU, gated multimodal unit; GRAD CNN, gradient regularized convolutional neural network; LSTM VGG-16, long short-term memory VGG-16; M, multi-module; NIN, network in network; ResMT, residual multi-task network; ResNet, residual neural network; ResNet101, residual neural network 101; ResNet50, residual neural network 50; S, single-module; SASG-GCN, self-attention similarity guided graph convolutional network; TCGA, The Cancer Genome Atlas; TCIA, The Cancer Imaging Archive; TDC-Distill, topology-preserving and discrepancy-eliminating contrastive distillation. Grading methods based on deep learning †, the network models that perform better each year for reference. The network models that perform better each year are bolded in the table for reference. Attention ResNet, attention residual network; AUC, area under the curve; BraTS, Brain Tumor Segmentation; CapsuleNet, capsule network; CNN, convolutional neural network; DLS-DARTS, double-learnable-stem differentiable architecture search; FCN, fully convolutional network; GAN, generative adversarial network; GMU, gated multimodal unit; GRAD CNN, gradient regularized convolutional neural network; LSTM VGG-16, long short-term memory VGG-16; M, multi-module; NIN, network in network; ResMT, residual multi-task network; ResNet, residual neural network; ResNet101, residual neural network 101; ResNet50, residual neural network 50; S, single-module; SASG-GCN, self-attention similarity guided graph convolutional network; TCGA, The Cancer Genome Atlas; TCIA, The Cancer Imaging Archive; TDC-Distill, topology-preserving and discrepancy-eliminating contrastive distillation. In this paper on glioma classification, CNNs and its variant networks perform more prominently compared to other networks, mainly due to their strong ability in automatically extracting hierarchical features from images, dealing with spatial invariance, and learning multilevel feature representations, and at the same time, the performance of the networks has been further improved by the introduction of innovative structures such as residual connectivity, dense connectivity, and so on, which enable them to better cope with medical complexities in image analysis. The ResMT model performs best due to its unique hybrid architecture design that combines the strengths of CNN and Transformer. Specifically, the spatial residual module (SRM) in the model captures the volumetric complexity of gliomas using 3D CNNs while augmenting the network with the pre-trained segmentation model, Swin UNETR, to boost performance without additional training. In addition, the multiplanar channel and spatial attention module (MCSA) refines the analysis by focusing on key features in multiple planes such as axial, coronal, and sagittal, which further enriches the model’s understanding of tumor features. The Transformer module, on the other hand, establishes a long-range relationship between planes and slices and efficiently solves the shortcomings of the CNNs in capturing the long-range dependencies between pixels, which highlights the hybrid CNN-Transformer model’s potential in MRI classification. Genotype classification The combination of tumor morphology and gene molecular (158) characteristics to treat individual patients is recognized by more and more surgeons and scholars. Diffuse gliomas are classified into grade 2–4 based on histopathology, but grade alone does not accurately reflect patient prognosis: even within the same pathological grade, patient prognosis will vary greatly. Recent advances in molecular biology have greatly improved our understanding of diffuse glioma. In 2016, WHO added molecular features to the histological classification of glioma for the first time. Isocitrate Dehydrogenase (IDH) and 1p19q became the key genetic markers for glioma molecular typing. WHO-CNS5 focuses on incorporating features such as genes, molecules, and alterations in signaling pathways into the classification system. IDH1 mutant, 1p/19q co-deleted tumors are mostly oligodendroglia histology: these tumors have the best prognosis/ longer survival. IDH1 mutant, 1p/19q co-deleted tumors are mostly oligodendroglia histology: these tumors have the best prognosis/ longer survival. IDH1 mutant, 1p/19q non co-deleted tumors that are mostly astrocytic morphology: these tumors have an intermediate survival. IDH1 mutant, 1p/19q non co-deleted tumors that are mostly astrocytic morphology: these tumors have an intermediate survival. IDH wild-type, 1p/19q non co-deleted tumors: these have a poor prognosis/short survival. IDH wild-type, 1p/19q non co-deleted tumors: these have a poor prognosis/short survival. IDH IDH is a significant genotype in the diagnosis and treatment of glioma. Non-invasive detection of IDH genotypes before surgery plays an important role in the treatment plan and survival prediction of patients. Lu et al. (38) found that pyramid dilated convolution is able to increase the receptive field of the original network efficiently. Consequently, the researchers incorporated pyramid dilated convolution into the ResNet framework to discern mutations from wild-type gliomas. Li et al. (37) integrated deep learning with radiomics to prognosticate the IDH status. Employing a modified CNN structure, the researchers undertook glioma segmentation and feature extraction from various MR imaging modalities, achieving a validation dataset accuracy of 95%. Choi et al. (159) also endeavored to predict the IDH mutation status of gliomas by leveraging deep learning and radiomics. Their approach involved a CNN-based segmentation model to capture radiomics features from the three-dimensional tumor shape, coupled with a CNN-based classifier for IDH status prediction. This hybrid model demonstrated an accuracy of 93.8%. Ahmad et al. (160) implemented an innovative CNN model without segmentation operations, focusing on pinpointing the most discriminative regions. Their investigation revealed the significance of the tumoral area in the majority of IDH mutant cases, while peritumoral edema was identified as influential in wild-type tumor patients. To address the challenges of insufficient and inefficient datasets, a novel paired GAN architecture (161) can be used to mitigate overfitting and improve performance. Using an open TCGA dataset of 167 patients, they observed a performance improvement of 2.57% when incorporating a mixture of real and augmented data. Furthermore, they proposed a two-stage-course training strategy (162) involving GAN-augmented MRIs followed by real MRIs, yielding a commendable test accuracy of 88.82%. Ai et al. (163) introduced the Three-Directional Attention Block Network (TDABNet) based on a 3D CNN and attention mechanism to capture additional information about gliomas in the third dimension. The AUC of IDH status prediction achieved an impressive 96.44%. Considering the relationship between IDH mutation and tumor angiogenesis in gliomas, Choi et al. (36) employed a LSTM model with an attention mechanism and dynamic susceptibility contrast (DSC) perfusion MRI to predict IDH status with an accuracy of 92.8%. Cheng (94) developed a hybrid CNN-Transformer to extract features from different MRIs and fuse the information into a semi-supervised learning approach, generating a substantial amount of unlabeled data. Experimental results demonstrated promising performance. To enhance IDH status prediction accuracy, Shi et al. (164) proposed an intra- and inter-modality fusion model, integrating MRI radiomics with deep learning features and utilizing an inter-modality regression model. Chang et al. (35) trained a residual CNN with each MR sequence (FLAIR, T2, T1 pre-contrast, and T1 post-contrast) and built a predictive model to determine IDH status, achieving an accuracy of 82.8%. Nalawade et al. (165) demonstrated the efficacy of a fully automated two-dimensional densely connected model trained exclusively on T2w MRI to predict IDH mutation status, achieving a mean classification accuracy of 90.5%. This suggests that multiple advanced methods and techniques can significantly improve the accuracy of preoperative noninvasive testing for IDH genotype prediction in gliomas, thus playing an important role in treatment planning and survival prediction for patients. This is achieved by incorporating pyramid expansion convolution into the ResNet framework, combining deep learning with radiomics, innovating CNN model design, applying a GAN architecture and a two-stage training strategy, introducing 3D CNNs and an attention mechanism, employing an LSTM model with an attention mechanism and DSC-perfused MRI, developing a hybrid CNN-Transformer model, proposing intra-modal and intermodal fusion models, and training 2D dense connectivity models for specific sequences, etc., to mine the feature information in MRI images from different perspectives and continuously optimize the prediction models. These methods are not only innovative in theory, but also have achieved remarkable results in practical applications, fully demonstrating the current research progress and great potential in the field of glioma IDH genotype prediction. 1p/19q The co-deletion of chromosome arms 1p/19q in LGG holds pivotal significance in formulating effective treatment strategies for patients. Akkus et al. (34) introduced a Multi-scale CNN with inherent self-learning capabilities, demonstrating proficient performance in predicting the status of 1p/19q using T1 and T2 images. Yogananda et al. (166) developed a T2-weighted image-only network (1p/19q-net), proficient in 1p/19q co-deletion status classification and concurrent glioma segmentation from T2-weighted images, utilizing 3D-Dense-UNets. This approach validates the efficacy of achieving high 1p/19q co-deletion classification accuracy solely using T2w MR images. Zhao et al. (167) implemented an efficient pipeline based on the U-net network to predict the status of 1p/19q from multi-modal MRI images. Comparative analysis with ResNet152 indicated higher sensitivity, specificity, and accuracy. Yan et al. (168) developed and validated a deep learning imaging signature (DLIS) for predicting the 1p/19q status in LGG patients, achieving an AUC of 0.983 for the testing dataset, showing the potential of deep learning as a noninvasive tool for 1p/19q status prediction. Jin et al. (169) devised the squeeze-and-excitation block DenseNet with weighted cross-entropy (SD-Net_WCE) for predicting the 1p/19q status, achieving an accuracy of 87.5% on data from 56 patients with 17,262 histological patch images. Kim et al. (170) conducted a comparative analysis of texture, topological, and CNN based image features. Matsui et al. (171) employed a deep learning model with MRI, PET, and CT, determining that the model achieved superior performance when the dataset included MRI, PET, and CT data. Ning et al. (172) devised a learning framework through a bi-directional mapping connecting the original imaging space and latent space, enabling the exploration of complementary information from multiple views. These methods are not only capable of extracting multi-scale features to enhance the recognition of multi-scale targets, but also efficiently perform image segmentation and feature extraction, such as the T2-weighted pure image network with 3D-Dense-UNets, which verifies the effectiveness of achieving high 1p/19q co-deletion classification accuracy using only T2w MR images. Using model optimization tools such as deep learning imaging signatures (DLIS) and squeeze-and-excitation block DenseNet with weighted cross-entropy (SDNet_WCE), as well as multimodal data fusion and feature comparison analysis, and bi-directional mapping learning frameworks, not only can we make full use of the complementary information of different modal data to provide a richer feature for model representation, but also further improves the accuracy and robustness of prediction. Multi-marker and integrated diagnosis WHO integrated diagnosis The WHO-CNS5 integrated diagnosis framework stratifies adult-type diffuse gliomas based on the defining molecular markers IDH mutation and 1p/19q codeletion status. Several deep learning studies have focused on predicting these class-defining markers from preoperative MRI. Chang et al. (173) employed a CNN to classify IDH mutation and 1p/19q codeletion status in 259 glioma patients, achieving accuracies of 94% and 92%, respectively. Buda et al. (174) utilized a deep learning model to extract genomic data including IDH mutation and 1p/19q co-deletion from 110 lower-grade glioma patients, finding an association between tumor shape features from MRI and glioma gene subtypes. Li et al. (175) investigated the interplay among glioma morphology, IDH1 genotypes, and 1p/19q chromosomes using a CNN to extract multi-modal glioma data from MRI, finding that IDH showed a stronger correlation with glioma MRI features compared to 1p/19q. Incekara et al. (176) developed a multi-task CNN for combined molecular subtyping, grading, and segmentation of glioma, achieving an IDH AUC of 0.90 and a 1p/19q co-deletion AUC of 0.85. Zhu et al. developed a three-classification machine learning model for the non-invasive prediction of molecular subtypes in diffuse glioma in a two-center study (177). These studies demonstrate that the defining molecular markers for WHO-CNS5 integrated classes can be predicted from routine MRI with clinically useful accuracy. Treatment-predictive markers MGMT promoter methylation status is a key treatment-predictive biomarker that guides alkylating chemotherapy decisions but is not used for WHO classification. Tupe-Waghmare et al. (178) employed a semi-supervised hierarchical multi-task model to concurrently predict IDH, MGMT, 1p/19q, and grade, achieving an average test accuracy of 82.35%. Nalawade (165) evaluated the impact of simulated motion artifacts on glioma marker classifier performance, including MGMT methylation, finding that motion artifacts directly influence classification accuracy while motion correction can restore and improve results. These findings highlight the importance of MGMT prediction as a treatment-predictive, rather than class-defining, marker. Prognostic and ancillary markers Additional molecular markers including Ki-67 proliferation index, TP53 mutation, and TERT promoter status serve supplementary prognostic roles but do not define WHO-CNS5 entities. Xu et al. (179) introduced a Vision Transformer model within a multitask classification framework to predict four different genotypes, including IDH, MGMT, Ki67, and P53 expression, demonstrating superior performance compared to T2-net, T1C-net, and TU-net baselines. Fukuma et al. (180) applied AlexNet to obtain CNN texture features for three genotypes incorporating TERT promoter mutations, classifying them with a linear support vector machine. Karami et al. (181) augmented conventional MRI with multi-shell diffusion MRI and utilized deep learning-based predictive models to discriminate three molecular subtypes, suggesting that advanced diffusion MRI provides complementary information for molecular prediction. These markers, while not class-defining, provide valuable supplementary prognostic information that complements the integrated diagnosis. Joint detection, segmentation and classification Cheng et al. (94) proposed a fully automated multimodal MRI-based multi-task learning framework for the joint tasks of glioma segmentation and IDH genotyping, demonstrating that a single network can effectively learn shared representations across the two tasks. An emerging paradigm integrates tumor detection, segmentation, and molecular marker prediction within unified end-to-end frameworks. Chakrabarty et al. (182) proposed a 2.5D hybrid multi-task CNN for joint prediction of IDH mutation and 1p/19q codeletion status from routine MRI, with multi-institutional external validation achieving AUCs of 0.931 (IDH) and 0.906 (1p/19q) on the external cohort. Chakrabarty et al. (183) developed a 3D CNN based on Mask R-CNN that jointly detects, segments, and predicts IDH mutation status from multi-modal MRI without requiring pre-segmented tumor masks. These joint approaches mitigate error propagation compared to sequential segmentation-then-classification pipelines and represent a promising direction toward fully automated glioma diagnostic systems. Utilization of datasets The frequency of usage of different databases in deep learning-based genotype classification of gliomas shows significant differences. In the survey of this paper, as shown in Figure 4, private databases top the list, followed by TCIA, and BraTS and TCGA databases are relatively less frequently used, although they are valuable in the field. Summary of the proportion of databases used for glioma genotype classification. NIMHANS, National Institute of Mental Health and Neuro Sciences; TCGA, The Cancer Genome Atlas; TCIA, The Cancer Imaging Archive. Private databases are highly favored for deep learning-based studies of glioma genotype classification, mainly due to their ability to provide unique, high-quality and diverse data resources. The data in these databases are often carefully selected and labeled in detail, enabling researchers to construct deep learning models with greater accuracy and generalization capabilities. In addition, the high novelty of data in private databases helps researchers explore new research directions and publish innovative results. At the same time, private databases also ensure the confidentiality of the data during the research process, avoiding the disclosure of the research results before publication, and allow researchers to customize the dataset according to their research needs, thus controlling the research process and direction more effectively. Many private datasets are unpublished synthetic datasets, and despite the advantages of private databases in these respects, researchers should also consider the importance of data sharing for the overall advancement of the scientific community. A compilation of literature on glioma genotype classification based on deep learning is presented in Table 5, categorized by IDH, 1p/19q, and multi-genotype. Table 5Glioma genotype classification based on deep learningRef.YearMarkerDatasetSampleModuleNetworkPerformanceEvaluation-protocol(34)20171p/19qPrivate159 scansMMulti-CNNAccuracy =87.7%External test(37)2017IDHPrivate151 patientsSCNNAUC =0.92External test(35)2018IDHPrivate, TCIA594 patientsSResNetAccuracy =89.1%External test(173)2018IDH, 1p/19q, MGMTTCIA259 patientsSCNN†IDH accuracy =94%, 1p/19q accuracy =92%, MGMT accuracy =83%External test(170)20191p/19qBraTS 2017167 scansSTDAAUC =0.684Validation(160)2019IDHNIMHANS71 patientsSCNN†Accuracy =86.7%External test(161)2019IDHTCGA167 patientsSGANAccuracy =81.03%External test(162)2020IDHTCGA167 patientsMGAN, U-NetAccuracy =88.82%External test(36)2019IDHPrivate463 patientsSAttention LSTMAccuracy =92.8%External test(180)2019IDH, TERTPrivate127 patientsSAlexNetMean Accuracy =63.1%External test(166)20201p/19qTCIA, TCGA368 casesMDense U-Nets†AUC =0.95External test(171)20201p/19qPrivate217 scansSResNetAccuracy =68.7%External test(38)2020IDHPrivate193 patientsSPyramid ResNetAccuracy =80.11%External test(175)2020IDH1 1p/19qPrivate360 patientsSXGBoostAccuracy =80%External test(184)2020IDH 1p/19q MGMTPrivate161 patientsSGAN1p/19q accuracy =74.81%, IDH accuracy =81.19%External test(169)20211p/19qPrivate323 casesSSDNetAccuracy =87.5%External test(172)2021IDH 1p/19qPrivate, TCGA, TCIA645 patientsMMulti-CNNIDH AUC =0.924, 1p/19q AUC =0.902External test(30)2021IDH 1p/19qBraTS 2019, Private, TCIA738 patientsSU-Net†TCIA IDH accuracy =94%, 1p/19q accuracy =82%, Private IDH accuracy =86%, 1p/19q accuracy =87%External test(178)2021IDH, MGMT, 1p/19qPrivate, TCIA307 casesMMulti-CNNMean accuracy =82.35%External test(168)20221p/19qPrivate, TCIA225 casesSDLISAUC =0.983External test(163)2022IDHBraTS 2019, TCGA235 patientsSTDABNetAUC =0.9644External test(164)2022IDHPrivate488 patientsSBRNNAccuracy =77%External test(94)2022IDHBraTS 2020439 patientsSU-Net, TransformerAccuracy =85.57%Internal test(179)2022IDH, MGMT, Ki67, P53Private188 patientsSU-Net, Transformer†IDH accuracy =98.5%, MGMT accuracy =98.35%, Ki67 accuracy =98.03%, P53 accuracy =98.58%External test(167)20231p/19qPrivate, TCGA150 casesSU-NetAccuracy =92.156%External test(176)2023IDH, 1p/19qPrivate1,748 patientsSFCNIDH AUC =0.90, 1p/19q AUC =0.85External test(185)2023IDHTCIA407 patientsSGNNAUC =0.962External test(186)2023CDKN2A/BTCGA, TCIA234 patientsMFN-Net†AUC =0.9704External test(187)2024IDH1Private315 patientsMMultiGeneNetAUC =0.886External test(188)2024IDHTCIA, Private759 patientsMMFEFNetAUC =0.8564External test(189)2024IDH, 1p/19qPrivate, TCIA424 casesS3D U-Nets†IDH AUC =0.96, 1p/19q AUC =0.961External test(190)2025IDH1, Ki67, MGMT, P53Private, TCIA555 casesMGDI-PMNet†IDH1 accuracy =98.31%, MGMT accuracy =97.96%, Ki67 accuracy =99.24%, P53 accuracy =98.54%External test(191)2025IDHPrivate402 patientsMResNet101AUC =0.98External test(192)20251p/19qPrivate, TCIA673 casesM1p/19qNETAUC =0.93External test(193)2025IDH, 1p/19qPrivate, TCGA, UCSF1,730 casesMGlioMTIDH TCGA AUC =0.915, IDH UCSF AUC =0.981, 1p/19q TCGA AUC =0.854, 1p/19q UCSF AUC =0.806External test†, the network models that perform better each year for reference. The network models that perform better each year are bolded in the table for reference. Attention LSTM, attention long short-term memory; AUC, area under the curve; BraTS, Brain Tumor Segmentation; BRNN, bidirectional recurrent neural network; CNN, convolutional neural network; 3D, three-dimensional; Dense U-Nets, densely connected U-nets; DLIS, deep learning imaging signature; FCN, fully convolutional network; FN-Net, fusion network; GAN, generative adversarial network; GDI-PMNet, gradient-aware dual-attention iteration truth calibration pyramid–mamba network; GlioMT, glioma multimodal transformer; GNN, graph neural network; M, multi-module; MFEFNet, multi-level feature exploration and fusion network; MGMT, O⁶-methylguanine-DNA methyltransferase; Multi-CNN, multi-branch convolutional neural network/multi-dimensional U-convolutional neural network; MultiGeneNet, multi-gene network; 1p/19qNET, 1p/19q codeletion prediction network; ResNet, residual neural network; ResNet101, residual neural network 101; S, single-module; SDNet, semantic decomposition network/spatial decomposition network; TCGA, The Cancer Genome Atlas; TCIA, The Cancer Imaging Archive; TDA, topological data analysis; TDABNet, three-directional attention block network; UCSF, University of California, San Francisco; XGBoost, extreme gradient boosting. Glioma genotype classification based on deep learning †, the network models that perform better each year for reference. The network models that perform better each year are bolded in the table for reference. Attention LSTM, attention long short-term memory; AUC, area under the curve; BraTS, Brain Tumor Segmentation; BRNN, bidirectional recurrent neural network; CNN, convolutional neural network; 3D, three-dimensional; Dense U-Nets, densely connected U-nets; DLIS, deep learning imaging signature; FCN, fully convolutional network; FN-Net, fusion network; GAN, generative adversarial network; GDI-PMNet, gradient-aware dual-attention iteration truth calibration pyramid–mamba network; GlioMT, glioma multimodal transformer; GNN, graph neural network; M, multi-module; MFEFNet, multi-level feature exploration and fusion network; MGMT, O⁶-methylguanine-DNA methyltransferase; Multi-CNN, multi-branch convolutional neural network/multi-dimensional U-convolutional neural network; MultiGeneNet, multi-gene network; 1p/19qNET, 1p/19q codeletion prediction network; ResNet, residual neural network; ResNet101, residual neural network 101; S, single-module; SDNet, semantic decomposition network/spatial decomposition network; TCGA, The Cancer Genome Atlas; TCIA, The Cancer Imaging Archive; TDA, topological data analysis; TDABNet, three-directional attention block network; UCSF, University of California, San Francisco; XGBoost, extreme gradient boosting. In this paper’s investigation on glioma genotype classification, CNN and U-Net and their variant networks perform well. U-net combined with Transformer performs the best, mainly due to its multi-task learning mechanism using hard parameter sharing, which is able to mine common features among different tasks while reducing the risk of overfitting and repetitive computation, and it is especially suited for small-sample image data. The model fuses T2WI, T1CWI and the combination of both T2 + T1CWI data to extract glioma features from multiple perspectives, providing a rich basis for prediction. It focuses on global information and effectively extracts features related to molecular expression by dividing the image into 2D blocks and utilizing the multi-head self-attention mechanism to capture the long-distance dependency between pixels. In addition, the Transformer structure inherits the advantages of the self-attention mechanism, dynamically adjusts the sensory field, avoids the dependence of traditional CNN on accurate segmentation, simplifies the preprocessing steps, and makes the model more focused on image feature learning. Risk of bias A substantial number of studies report sample sizes in scans or slices rather than patients. When data are split at the slice level, multiple correlated slices from the same tumor can appear in both training and test sets, causing optimistic performance estimates through data leakage. In addition, several studies combine BraTS with TCIA or TCGA without acknowledging that BraTS training cases are drawn from these same public repositories. This creates patient overlap across partitions, so the same individual may appear in both training and external-test sets, even when labeled as external validation. The overwhelming majority of entries rely on single-center private data or a single public source. Deep-learning models can exploit scanner-specific artifacts rather than true biological signal, producing high in-house AUCs that collapse under multi-site validation. Only a handful of studies test across multiple independent sites, and even these show marked performance gaps across centers, underscoring site confounding. Furthermore, IDH and 1p/19q status are strongly correlated with WHO grade and patient age, yet none of the entries explicitly mention grade-stratified or age-matched splitting. Models may therefore learn to classify grade or age rather than genotype. Class imbalance is another concern. IDH-wildtype gliomas outnumber IDH-mutant cases in most cohorts, and 1p/19q codeletion is present in only a minority of IDH-mutant tumors. Very high AUCs reported in small single-center studies are statistically suspect without imbalance correction or confidence intervals. Discussion Challenges In the context of advancing precision medicine as a pivotal trajectory in medical progress, the utilization of deep learning, particularly leveraging extensive datasets, has gained heightened importance. Deep learning has demonstrated notable accomplishments in glioma segmentation, grading and genotype classification. Despite these achievements, it is imperative to acknowledge that these deep learning-based applications still face challenges as follows. Data Despite an increasing willingness among hospitals to share datasets related to gliomas for applications in deep learning research, the existing datasets continue to face challenges such as insufficient quantity, limited diversity in sources, varying levels of data quality, annotation bias, data imbalance, and privacy protection. These issues significantly hinder the widespread application of glioma related research and also constrain the practical implementation of corresponding artificial intelligence products. In comparison to other fields, acquiring high-quality medical data is relatively challenging. Additionally, for glioma, due to subjective patient factors and objective aspects of medical conditions, not all cases undergo homogenized, standardized diagnostic and treatment procedures, coupled with comprehensive follow-up. The data accumulated by individual medical institutions is often inadequate to effectively train learning models, multicenter studies have emerged as a plausible solution. However, medical and health data between regions and medical institutions remain in information isolation, lacking a unified model for sharing data resources, thereby impeding the realization of large-scale, high-quality sample data acquisition of glioma across multiple centers. Like other tumor datasets, glioma data also encounters the challenge of data imbalance. From a data perspective, techniques such as data augmentation, data resampling, and dynamic sampling can be employed. From an algorithmic standpoint, specific loss functions can be adopted to mitigate the adverse effects of data imbalance. Upon acquiring a sufficient dataset, the provision of well-annotated extensive sample datasets is necessary for training models. If the quality of annotated glioma data fails to meet standards, the deep learning model cannot be effectively trained through accurate input, resulting in incorrect model outputs. Models While deep learning models have advanced rapidly in recent years, classical network architectures, such as U-Net, typical CNNs and ResNet, remain the mainstream backbones chosen by researchers. The structural modification of the U-Net remains a focal point in current research within the field of glioma segmentation. In addition, techniques such as skip connections, attention mechanisms, and transfer learning are effectively integrated with U-Net to achieve improved segmentation, grading, and classification accuracy. Beyond data augmentation methods, from the perspective of model, researchers are increasingly exploring approaches that leverage limited datasets, employing techniques such as transfer learning and meta-learning. These methods aim to optimize model performance even when confronted with a scarcity of labeled medical images, thereby enhancing the robustness and generalization capability of the models in the context of segmentation or classification task. Also, the investigation into model interpretability, particularly within the realm of evidence-based medicine, is always a significant challenge. In the domain of medical imaging, interpretability of a model can assist physicians in discerning the features upon which the model bases its diagnostic decisions, thereby aiding healthcare professionals in better understanding the judgments made by the model. In fact, the attention mechanism is also one of the tools that helps explain the model. The attention mechanism can help the model pay more attention to important areas when processing images. By introducing the attention mechanism into the medical imaging deep learning model, the model can automatically learn areas that are more important for diagnostic results and focus more attention on these areas. In this way, doctors can understand the model’s decision-making basis by observing the model’s attention area. Applications Meeting the requirements of diverse application scenarios with a general model proves to be a formidable challenge due to the intricate and diffusive nature of glioma structures. The inherent characteristics include low contrast, uncertain morphologies, and limited prior knowledge about the tumors. Furthermore, variations in the size, shape, location, and tumor subtype among different patients contribute to the complexity. Meanwhile, certain tumors exhibit similarities to normal brain tissues, leading to indistinct boundaries that are even challenging for manual differentiation. While significant progress has been made in the segmentation, grading, and genotype classification of gliomas through numerous research efforts, the stringent and critical nature of the medical field demands utmost precision. Even with notable advancements in algorithms, the deployment of models for practical clinical use remains quite limited, as the field of medicine prioritizes patient safety, and any slight algorithmic error must not jeopardize patient lives. Clinical deployment and regulatory considerations Despite the substantial research progress documented in this survey, a significant gap persists between the development of high-performing deep learning algorithms for glioma analysis and their routine deployment in clinical practice. The translation of AI models from research environments to clinical workflows requires navigation of complex regulatory pathways. In the United States, the Food and Drug Administration (FDA) regulates AI-based medical devices under either the 510(k) premarket notification pathway or the De Novo classification process. In Europe, the CE marking process under the Medical Device Regulation (MDR) requires conformity assessment for AI diagnostic tools. In China, the National Medical Products Administration (NMPA) has established dedicated guidance for AI medical devices, requiring prospective clinical trials and multi-center validation for class III devices. To date, only a limited number of prospective clinical trials have evaluated deep learning tools specifically for glioma segmentation or molecular marker prediction. Most studies reviewed in this survey rely on retrospective evaluation on curated datasets, which may not reflect the heterogeneity of real-world clinical imaging. Practical deployment barriers extend beyond regulatory approval to encompass scanner variability across vendors and field strengths, integration with existing Picture Archiving and Communication Systems (PACS), clinician trust and acceptance of AI-generated outputs, and unresolved medico-legal questions regarding liability for algorithmic errors. Future research directions Drawing on the methodological findings summarized in Tables 3-5 and the challenges identified in the preceding discussion, this section outlines priority directions for future research in deep learning-based glioma analysis. Prospective multi-center validation and external testing. The overwhelming majority of studies reviewed in this survey rely on internal validation or BraTS challenge test sets that may overlap with training data across challenge editions. Future work must prioritize prospective, multi-center validation on cohorts acquired independently from the training data, spanning multiple scanner vendors, field strengths, and imaging protocols. Prospective multi-center validation and external testing. The overwhelming majority of studies reviewed in this survey rely on internal validation or BraTS challenge test sets that may overlap with training data across challenge editions. Future work must prioritize prospective, multi-center validation on cohorts acquired independently from the training data, spanning multiple scanner vendors, field strengths, and imaging protocols. Standardized evaluation protocols and reproducibility. The segmentation, grading, and genotype classification literatures employ heterogeneous evaluation designs—including k-fold cross-validation, single hold-out testing, challenge leaderboard submissions, and external cohort evaluation—that are not directly comparable even when the same metric is reported. The field would benefit from standardized benchmarking protocols and public release of source code and trained model weights. Standardized evaluation protocols and reproducibility. The segmentation, grading, and genotype classification literatures employ heterogeneous evaluation designs—including k-fold cross-validation, single hold-out testing, challenge leaderboard submissions, and external cohort evaluation—that are not directly comparable even when the same metric is reported. The field would benefit from standardized benchmarking protocols and public release of source code and trained model weights. Harmonized train/validation/test reporting. Tables 3-5 reveal that most studies do not clearly distinguish between validation-set performance, internal test-set performance, and challenge test-set performance. Future publications should explicitly report results on all three splits. Harmonized train/validation/test reporting. Tables 3-5 reveal that most studies do not clearly distinguish between validation-set performance, internal test-set performance, and challenge test-set performance. Future publications should explicitly report results on all three splits. Missing-sequence and limited-modality evaluation. Clinical practice frequently encounters MRI scans with one or more unavailable sequences. Future methods should incorporate missing-modality handling as a core design requirement. Missing-sequence and limited-modality evaluation. Clinical practice frequently encounters MRI scans with one or more unavailable sequences. Future methods should incorporate missing-modality handling as a core design requirement. Calibration and uncertainty quantification. High accuracy alone is insufficient for clinical deployment; models must also provide well-calibrated probability estimates and meaningful uncertainty quantification. Calibration and uncertainty quantification. High accuracy alone is insufficient for clinical deployment; models must also provide well-calibrated probability estimates and meaningful uncertainty quantification. Integrated segmentation and molecular prediction pipelines. Joint detection, segmentation, and molecular classification frameworks that operate directly on raw MRI without requiring pre-segmented inputs represent a promising direction toward end-to-end diagnostic systems. Integrated segmentation and molecular prediction pipelines. Joint detection, segmentation, and molecular classification frameworks that operate directly on raw MRI without requiring pre-segmented inputs represent a promising direction toward end-to-end diagnostic systems. Robustness to scanner, site, and distribution shifts. Variations in MRI scanner manufacturer, field strength, and pulse sequence parameters introduce substantial distribution shifts. Future research should systematically evaluate cross-scanner and cross-site generalization. Robustness to scanner, site, and distribution shifts. Variations in MRI scanner manufacturer, field strength, and pulse sequence parameters introduce substantial distribution shifts. Future research should systematically evaluate cross-scanner and cross-site generalization. Foundation model adaptation for glioma. Vision Transformers, SAM, MedSAM, and related foundation models offer powerful pre-trained representations that can be adapted to glioma-specific tasks through parameter-efficient fine-tuning, potentially reducing reliance on large labeled datasets. Foundation model adaptation for glioma. Vision Transformers, SAM, MedSAM, and related foundation models offer powerful pre-trained representations that can be adapted to glioma-specific tasks through parameter-efficient fine-tuning, potentially reducing reliance on large labeled datasets. Conclusions This paper offers a comprehensive review of deep learning-based techniques tailored for glioma segmentation, grading, and genotype classification. The past several years have witnessed diverse approaches and network models applied to different glioma tasks. The collective insights from these studies contribute significantly to the evolving landscape of glioma diagnosis, paving the way for more effective and efficient healthcare practices. Looking ahead, it is essential to acknowledge and address the common challenges encountered in this field. Future research directions should be shaped by a commitment to overcoming these challenges, thereby fostering the continued advancement of deep learning applications in glioma diagnosis. By fostering collaboration, refining existing methodologies, and exploring innovative approaches, the medical community can harness the full potential of deep learning to enhance diagnostic precision and ultimately improve patient outcomes in the realm of glioma management. Supplementary The article’s supplementary files as