A transformer-based multimodal deep learning model for preoperative prediction of Ki-67 expression level in glioma Authors: Liu Liu, Fangfang Hu, Minquan Lai, Xianjing Zhao, Pinggui Lei, Zhenwei Yao, Bing Fan Published: 2026 ABSTRACT AbstractPurposeThe aim of this study is to develop a deep learning model using preoperative multimodal MR data to predict the Ki-67 expression level of glioma and externally validate the predictive performance of the model.MethodsThis study retrospectively collected the clinical and imaging data from 421 patients with grade 2–4 gliomas who underwent surgical resection or biopsy and were pathologically diagnosed in two hospitals between January 2020 and December 2024. The 421 patients were divided into a training set (N = 217), an internal validation set (N = 94), and an external validation set (N = 110). Then, the tumor margins were delineated on contrast-enhanced T1- weighted imaging (CE-T1WI) and contrast-enhanced T2-fluid-attenuated inversion recovery (CE-T2FLAIR) to obtain the three-dimensional region of interest (3D ROI) of the tumor. Three vision transformer (ViT) models based on CE-T1WI, CE-T2FLAIR, and CE-T1WI + CE-T2FLAIR (CE-T1WI_ViT, CE-T2FLAIR_ViT and Combined_ViT) were constructed respectively. The predictive performance of the models was evaluated by the area under the receiver operating characteristic curve (AUC). Finally, to further assess the predictive performance of our transformer model, we trained and tested three convolutional neural network (CNN) models (ShuffleNet, ResNet50, and DenseNet121) on the same dataset and compared our trained Combined_ViT model with these three CNN models.ResultsThe three models, including CE-T1WI_ViT, CE-T2FLAIR_ViT, and Combined_ViT, demonstrated high predictive accuracy for Ki-67 level in grade 2–4 gliomas, with AUC values of 0.859 (95% confidence interval [CI], 0.792–0.926), 0.825 (95%CI, 0.745–0.906), and 0.922 (95%CI, 0.868–0.976), respectively. Among the three models, the Combined_ViT model achieved the highest predictive accuracy. Furthermore, the predictive performance of the Combined_ViT model exceeded that of the three CNN models (ShuffleNet, ResNet50, and DenseNet121), with AUC values of 0.897 (95%CI, 0.839–0.954), 0.905 (95%CI, 0.843–0.966), and 0.913 (95%CI, 0.861–0.965) respectively.ConclusionsThe deep learning models based on ViT can effectively predict the Ki-67 expression level of glioma, and are a feasible alternative to CNN models.Clinical trial numberNot applicable.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12880-026-02492-z.Keywords: Glioma, Vision transformer, Convolutional neural network, Ki-67 index Introduction Glioma arises from neural glial cells in the brain and is the most common primary malignant tumor in the central nervous system, characterized by high disability and mortality rates [1]. Ki-67, a nuclear protein antigen mainly encoded by the MKI67 gene, plays a critical role in cell proliferation. Its expression is markedly increased during the G1, S, G2, and M phases of the cell cycle, while absent in the resting phase [2]. In normal brain tissue, Ki-67 level is minimal. However, elevated Ki-67 level is strongly correlated with poor tumor prognosis. A higher Ki-67 index indicates greater malignancy, worse prognosis, and reduced progression-free survival and overall survival rates [3]. Thus, accurately predicting Ki-67 expression in glioma is of great clinical significance for the formulation of individualized treatment plans and the prognosis assessment for patients. However, Ki-67 detection necessitates immunohistochemical analysis after tumor resection or biopsy. But these invasive methods are costly, carry certain risks, and may be intolerable for some patients. Therefore, developing a non-invasive tool to evaluate Ki-67 expression in glioma patients is of considerable clinical importance. In recent years, with the continued advancement of deep learning(DL) technology, the non-invasive predictive model based on DL has gradually been applied to the preoperative grading (staging), molecular subtype prediction, treatment efficacy evaluation and prognosis assessment across various types of tumors [4–6]. CNNs are currently among the most widely used DL architectures. By simulating the way the human brain processes visual information, CNNs can automatically extract features from medical images that are not easily quantifiable by the human eye, such as tumor shape, edge, texture, and they can also detect subtle differences between image types, enabling tasks such as automated lesion detection, segmentation, and classification [7–8]. Besides, CNNs leverage the parameter-sharing mechanism to drastically reduce the overall model parameter count, which not only facilitates efficient training but also ensures stable performance in small-sample scenarios commonly encountered in the medical field [9]. However, due to their inherently local receptive fields, CNNs mainly capture information from neighboring pixels and lack the ability to establish global dependencies [10]. Another DL architecture, the ViTs, have emerged as a promising alternative. Since 2020, ViTs have been introduced into the field of computer vision, leveraging the self-attention mechanism originally developed for natural language processing to overcome the locality limitations of CNNs [11]. Compared with CNNs, ViTs offer higher accuracy and stronger generalization capabilities, holding significant research and clinical value for tumor image segmentation, molecular subtype predictive, and prognosis assessment [12–14]. With the continuous deepening of research, some scholars have even proposed hybrid architectures that integrate CNNs and ViTs. Such frameworks combine the superior local feature extraction capability of CNNs with the powerful global context modeling ability of ViTs, thereby achieving feature complementarity and synergistic interpretability [15, 16]. Up to now, most current DL models for glioma have been based on conventional MR sequences such as T2WI, T2FLAIR, and CE-T1WI, while other sequences, including Diffusion Weighted Imaging(DWI), Perfusion Weighted Imaging(PWI) and CE-T2FLAIR are rarely used in DL [17]. To date, no studies have established DL model based on CE-T2FLAIR. However, in recent years, with more in-depth research on CE-T2FLAIR, neuroimaging scholars have identified several advantages over CE-T1WI, such as improved detection and differentiation of brain micro-metastases and meningeal lesions, as well as the ability to distinguish radiation necrosis from tumor progression [18, 19]. Consequently, CE-T2FLAIR has been increasingly applied in clinical practice. One of our studies aims to construct a DL model for predicting the Ki-67 expression level of glioma based on CE-T2FLAIR and to verify the prediction performance of the model on an external validation cohort(EVC). Materials and methods Data collection This study retrospectively collected datasets from two centers and was conducted in accordance with the Declaration of Helsinki. The study was approved by the Medical Ethics Committee of Jiangxi Provincial People’s Hospital (No. 2024 [68]), and the requirement for informed consent was waived. Huashan Hospital was also notified and approved the study. The detailed patient enrollment procedure is shown in Fig. 1. Center 1 (Huashan Hospital Affiliated to Fudan University) included 362 patients with grade 2–4 adult diffuse gliomas from April 2020 to June 2023, while Center 2 (Jiangxi Provincial People’s Hospital) included 133 patients with grade 2–4 adult diffuse gliomas from June 2021 to December 2024. All patients were diagnosed with grade 2–4 adult diffuse gliomas through surgical resection or biopsy. And all subtypes of these adult diffuse gliomas were classified with reference to the 2021 WHO Classification of Tumours of the Central Nervous System. Clinical, pathological, and preoperative imaging data were collected for all patients. The inclusion criteria for this study were as follows: (1) pathological confirmation of adult diffuse gliomas (grade 2–4) with known Ki-67 expression level; (2) brain MRI performed within two weeks before surgery. The exclusion criteria were: (1) unsatisfactory image quality due to various artifacts; (2) lack of CE-T1WI or CE-T2FLAIR before surgery. In accordance with previous research methods [20], Ki-67 expression level was classified into four categories: 0–4% as grade 0, 5%-24% as grade 1, 25%-49% as grade 2, and over 50% as grade 3. For this study, Ki-67 expression level was further categorized into a low-expression group (grade 0–1) and a high-expression group (grade 2–3). In total, 421 patients with WHO grade 2–4 adult diffuse gliomas were included. The clinical and tumor characteristics of patients in each center are presented in Table 1. The training cohort included 126 patients with low Ki-67 expression and 91 patients with high Ki-67 expression, indicating a mild class imbalance. To reduce potential bias toward the majority class, the synthetic minority over-sampling technique (SMOTE) was applied to the training set. SMOTE was used to generate synthetic samples for the minority high Ki-67 class during model training. Flow chart of the patient inclusion criteria Table 1Clinical and tumor characteristics of patients in cohortsClinical and tumor characteristicsTraining cohort(N = 217)Internal validation cohort (N = 94)External validation cohort (N = 110)Ki67<25%(N = 126) Ki67 ≥ 25%(N = 91) P-valueKi67<25%(N = 56) Ki67 ≥ 25%(N = 38) P-valueKi67<25%(N = 55) Ki67 ≥ 25%(N = 55) P-value Age 47.103 ± 15.39054.527 ± 14.655< 0.00145.464 ± 13.75548.211 ± 13.5590.34250.891 ± 14.58856.909 ± 13.1720.028 Gender 0.3550.5320.443 Female59(46.825)36(39.560)27(48.214)15(39.474)27(49.091)22(40.000) Male67(53.175)55(60.440)29(51.786)23(60.526)28(50.909)33(60.000) Tumor number 0.0040.1720.458 Single111(88.095)65(71.429)51(91.071)30(78.947)47(85.455)43(78.182) Multiple15(11.905)26(28.571)5(8.929)8(21.053)8(14.545)12(21.818) Largest diameters(cm) 0.5541.00.336 ≤ 5 cm70(55.556)46(50.549)27(48.214)19(50.000)27(49.091)21(38.182) >5 cm56(44.444)45(49.451)29(51.786)19(50.000)28(50.909)34(61.818) Tumors border 0.0060.910.142 Clear78(61.905)73(80.220)32(57.143)23(60.526)35(63.636)43(78.182) Blurred48(38.095)18(19.780)24(42.857)15(39.474)20(36.364)12(21.818) WHO grade < 0.001< 0.001< 0.001 249(38.889)0(0.000)16(28.571)1(2.632)28(50.909)0(0.000) 319(15.079)4(4.396)13(23.214)4(10.526)11(20.000)7(12.727) 458(46.032)87(95.604)27(48.214)33(86.842)16(29.091)48(87.273) Subtypes of adult-type diffuse gliomas Astrocytoma, IDH-mutant35(27.778)5(6.316)14(25.000)6(15.789)17(30.909)4(7.273)Oligodendroglio-ma36(28.571)2(3.158)16(28.571)1(2.632)10(18.182)1(1.818)Glioblastoma, IDH-wildtype55(43.651)84(90.526)26(46.429)31(81.579)28(50.909)50(90.909) Clinical and tumor characteristics of patients in cohorts MRI Imaging acquisition The brain MR images of all patients were obtained using Siemens 3.0 T and GE 3.0 T scanners. The magnetic resonance sequences included at minimum: (a) axial CE-TIWI; (b) axial CE-T2FLAIR. Detailed MR sequence parameters of the two centers are provided in Supplementary Material 1. Clinical data Clinical data included age (≤ 50 years, > 50 years), gender, and imaging features such as tumor numbers (solitary, multiple), tumor boundary (blurred, clear), and maximum tumor diameter (≤ 5 cm, > 5 cm). These characteristics are summarized in Table 1. Imaging data preprocessing The data from Center 1 were randomly divided into a training set and an internal validation cohort(IVC) in a 70%: 30% ratio, while the data from Center 2 were used as the EVC. To mitigate sampling bias and rigorously evaluate the model’s generalization performance, we implemented a 5-fold cross-validation only on the training cohort of Center (1) The internal validation cohort was used for preliminary verification of the model performance after cross-validation optimization, while the external validation cohort was used for the final evaluation of the model’s generalization ability. Prior to segmentation, all images underwent preprocessing. The original DICOM images from Centers 1 and 2 were first converted to the nii. gz format. Subsequently, the MR images with a 1 mm slice thickness from Center 1 were resampled to a 5 mm slice thickness to alleviate the extreme time and labor burden of manual glioma segmentation and match the slice thickness of the MR images from Center (2) Detailed implementation procedures, as well as the quantitative validation assessing the impact of downsampling on tumor boundaries and information loss, are provided in Supplementary Material 2. These findings confirm that clinically relevant tumor boundary information was well-preserved despite downsampling. Other preprocessing steps included image normalization, bias field correction and data augmentation techniques, with detailed implementations presented in Supplementary Material (3) Specifically, the preprocessing steps, including normalization, resampling, and data augmentation, were only fitted on the training set, and then applied to the validation and test sets. No information from the validation or test sets was used to determine preprocessing parameters. Finally, to reduce the potential bias caused by different scanner vendors, histogram matching was performed to standardize the intensity distribution across all images from different centers. In addition, z-score normalization was applied on a per-center basis to further minimize scanner-specific effects. Image segmentation Two radiologists with more than five years of experience in neuroimaging independently performed a manual delineation of three-dimensional regions of interest (3D ROIs) slice by slice and blinded to all clinical and pathological outcomes on the axial CE-T2FLAIR and CE-T1WI images. The delineation was carried out using the open-source software 3D Slicer 4.11.0 (https://www.slicer.org/). ROIs were delineated on CE-T1WI and CE-T2FLAIR sequences, encompassing areas such as tumor parenchyma, intratumoral hemorrhage, necrosis, and cystic liquefaction (Fig. 2). To minimize segmentation errors and ensure reliable feature extraction, other sequence images were also referenced during the delineation process. Inter-rater segmentation consistency was evaluated using the Dice coefficient. For segmentations with a Dice coefficient below 0. 95 or in cases of disagreement, the results were reviewed and adjusted by an experienced expert to ensure accuracy and high consistency of the final segmentation. The mean inter-rater Dice score for the full dataset is 0.96 ± 0.03 (mean ± standard deviation). Tumor segmentation. a and b show the tumor on CE-T1WI and CE-T2FLAIR. c and d show the segmentation of the tumor on CE-TIWI and CE-T2FLAIR. e shows the 3D shape of the segmented tumor Development of the ViT network This research introduces a novel 3D ViT architecture designed to accurately predict the Ki-67 proliferation index in gliomas. The schematic diagram of the model is shown in Fig. 3. The proposed model employs a distinctive 3D patching strategy, dividing each ROI into 16 × 16 voxel patches in the axial plane (XY) and stacking five consecutive slices per patch along the sagittal axis (Z). This approach generates a total of 2048 volumetric patches per ROI. Each patch is then linearly projected into a 768-dimensional feature vector and combined with 3D positional encodings incorporating X, Y, and Z coordinates to effectively capture spatial information. To enhance generalization, a dropout rate of 0.2 is applied to the embedding layers. In addition, a learnable token is prepended to the sequence for classification purposes. The core architecture of the model consists of six stacked transformer encoder layers. Each encoder layer includes a 12-head self-attention mechanism (with a head dimension of 64), a 2048-dimensional multi-layer perceptron (MLP), layer normalization with residual connections, and an attention dropout rate of 0.2. Model training hyperparameters, hardware environment, and relevant implementation details are provided in Supplementary Material 4. Schematic diagram illustrating the details of the ViT model used in this study Development of the CNN network To provide a robust comparative evaluation of predictive performance against the proposed ViT model, a comprehensive analysis was conducted using three well-established CNN architectures: ShuffleNet, ResNet50, and DenseNet121. These models were specifically adapted to handle 3D volumetric data, ensuring compatibility with the input requirements of the MRI dataset used in this study. Each of the selected CNN models was fine-tuned to optimize its performance in predicting the Ki-67 proliferation index in gliomas. The original fully connected top layers of these models, designed for different classification tasks and datasets, were removed. Instead, new top layers were constructed and trained from scratch to better capture the specific characteristics of glioma MRI data. This approach enabled the models to learn more relevant features directly from the dataset, thereby enhancing their predictive performance. The data preprocessing pipeline for these CNN models was aligned with that of the ViT model to ensure a fair and unbiased comparison. CNN-specific training hyperparameters are detailed in Supplementary Material 5. In addition, hyperparameter tuning was performed fairly across all architectures, ensuring objective performance comparisons. Statistical analysis Statistical analysis were conducted in SPSS(version 28.0) and R(version 4.2.3) with the package pROC(version 1.18.0). Measurement data were expressed as x ± s, and comparisons between groups were conducted using the independent sample t-test. Count data were presented as case numbers, and group comparisons were performed using the χ2 test. P value less than 0.05 was considered to indicate statistical significance. Result Clinical results The clinical and tumor imaging characteristics of 421 patients are shown in Table 1. There were no statistically significant differences in gender, numbers of tumors, tumor boundary, or maximum tumor diameter between the low Ki-67 expression group and the high Ki-67 expression group. Statistically significant differences were observed in tumor WHO grade (P < 0.05). Higher WHO grade was associated with higher Ki-67 expression in glioma. However, in the training set and EVC, there were statistically significant differences in age between the low and high Ki-67 expression groups(P < 0.05), whereas in the IVC, no statistically significant differences in age were observed between the two groups (P > 0.05). In addition, the results also indicated that there was a significant correlation between the molecular subtype of adult diffuse glioma and Ki-67 expression level in each dataset: IDH-mutant astrocytoma and oligodendroglioma were mainly distributed in the Ki-67 low-expression group, while IDH-wildtype glioblastoma was mainly distributed in the Ki-67 high-expression group. This suggested that the molecular subtype of adult diffuse glioma is closely associated with Ki-67 expression. IDH-mutant astrocytomas and oligodendrogliomas, which correspond to lower-grade and relatively indolent glioma entities, predominantly exhibit low Ki-67 expression, consistent with their slower proliferation and more favorable clinical behavior. In contrast, IDH-wildtype glioblastomas, which represent highly malignant and aggressive tumors, are strongly associated with high Ki-67 expression, reflecting their markedly elevated proliferative potential and aggressive biological behavior. Moreover, among the three subtypes of adult diffuse gliomas, glioblastoma exhibited the widest range of Ki-67 expression levels (from 4% to 80%), which reflects the remarkable intrinsic heterogeneity of this tumor entity. Establishment and validation of DL models All three DL models, CE-T1WI_ViT, CE-T2FLAIR_ViT, and Combined_ViT, performed relatively well in predicting the Ki-67 expression level of glioma in both the IVC and EVC. The ROC curves of the three models in the IVC and EVC are presented in Fig. 4, and the AUC, 95% CI, accuracy, sensitivity, specificity, positive prediction value(PPV), negative prediction value(NPV), and other performance metrics are summarized in Table 2. The decision curves of the three ViT models for predicting Ki-67 expression level in glioma are shown in Fig. 5. The curves demonstrate that, within a reasonable threshold probability, Combined_ViT offers a greater net benefit in distinguishing Ki-67 expression levels. The calibration curves of the three ViT models are presented in Fig. 6. The curves indicate that, in both the IVC and EVC, Combined_ViT aligns more closely with the diagonal, reflecting superior calibration compared to the other two ViT model. ROC of different models. (a) prediction results in IVC, (b) prediction results in EVC Table 2The performance of models based on ViT in predicting Ki-67 expression levelCohortModelAUCAUC 95%CIAccSenSpePPVNPVRecallF1TrainingCE-T2FLAIR_ViT0.9000.8534–0.94670.8760.8130.9210.8810.8720.8130.846TrainingCE-T1WI_ViT0.9160.8770–0.95500.8660.9010.8410.8040.9220.9010.850TrainingCombined_ViT0.9910.9781–1.00000.9770.9890.9680.9570.9920.9890.973IVCCE-T2FLAIR_ViT0.8290.7455–0.91240.7980.9210.7140.6860.9300.9210.787IVCCE-T1WI_ViT0.8370.7541–0.91930.7870.9470.6790.6670.9500.9470.783IVCCombined_ViT0.9440.8937–0.99500.9360.9470.9290.9000.9630.9470.923EVCCE-T2FLAIR_ViT0.8250.7452–0.90570.8090.9090.7090.7580.8860.9090.826EVCCE-T1WI_ViT0.8590.7922–0.92620.8000.9450.6550.7320.9230.9450.825EVCCombined_ViT0.9220.8682–0.97580.8910.9450.8360.8520.9390.9450.897Abbreviations: IVC, internal validation cohort; EVC, external validation cohort; Acc, accuracy; Sen, sensitivity; Spe, specificity; PPV, positive prediction value; NPV, negative prediction value The performance of models based on ViT in predicting Ki-67 expression level Abbreviations: IVC, internal validation cohort; EVC, external validation cohort; Acc, accuracy; Sen, sensitivity; Spe, specificity; PPV, positive prediction value; NPV, negative prediction value DCA of different models. (a) prediction results in IVC, (b) prediction results in EVC The calibration curve of different models. (a) prediction results in IVC, (b) prediction results in EVC All three CNN models demonstrated relatively good predictive performance for the Ki-67 expression level of glioma in both the IVC and EVC. The ROC curves of the three models in the IVC and EVC are shown in Fig. 7, and the values of AUC, 95% CI, accuracy, sensitivity, specificity, PPV, NPV, and other metrics for the three models are presented in Table 3. The DeLong test of ViT and CNN models are shown in Table 4. According to the DeLong test, among ViT models, the Combined_ViT model demonstrated the best predictive performance, while no significant differences in predictive performance were observed among these three CNN models. Finally, we compared these three CNN models with the Combined_ViT model. The results indicated that the predictive performance of the Combined_ViT model was slightly superior to that of the three CNN model. The decision curves of the three CNN models and the Combined_ViT model in the IVC and EVC for predicting the Ki-67 expression level of glioma are shown in Fig. 8. The decision curves demonstrate that, within a reasonable threshold probability range, Combined_ViT provides a greater net benefit in differentiating the Ki-67 expression level of glioma. The calibration curves of the three CNN models and the Combined_ViT model are presented in Fig. 9. The calibration curves indicate that in both the IVC and EVC, Combined_ViT exhibits better diagonal fits compared with the three CNN models. ROC of different models. (a) prediction results in IVC, (b) prediction results in EVC Table 3A comparison of the predictive performance between the ViT-based combined model and three combined models based on CNNCohortModelAUCAUC 95%CIAccSenSpePPVNPVRecallF1TrainingCombined_ShuffleNet0.9570.9282–0.98520.9260.8680.9680.9520.9100.8680.908TrainingCombined_ResNet500.9670.9409–0.99230.9350.9010.9600.9430.9310.9010.921TrainingCombined_DenseNet1210.9710.9517–0.98990.9170.8790.9440.9200.9150.8790.899TrainingCombined_ViT0.9910.9781–1.00000.9770.9890.9680.9570.9920.9890.973IVCCombined_ShuffleNet0.9150.8575–0.97290.8830.7111.0001.0000.8360.7110.831IVCCombined_ResNet500.9220.8577–0.98680.9150.9740.8750.8410.9800.9740.902IVCCombined_DenseNet1210.9270.8601–0.99430.9260.9470.9110.8780.9620.9470.911IVCCombined_ViT0.9440.8937–0.99500.9360.9470.9290.9000.9630.9470.923EVCCombined_ShuffleNet0.8970.8392–0.95420.8360.8000.8730.8630.8140.8000.830EVCCombined_ResNet500.9050.8432–0.96640.9001.0000.8000.8331.0001.0000.909EVCCombined_DenseNet1210.9130.8614–0.96510.8730.9820.7640.8060.9770.9820.885EVCCombined_ViT0.9220.8682–0.97580.8910.9450.8360.8520.9390.9450.897Abbreviations: IVC, internal validation cohort; EVC, external validation cohort; Acc, accuracy; Sen, sensitivity; Spe, specificity; PPV, positive prediction value; NPV, negative prediction value A comparison of the predictive performance between the ViT-based combined model and three combined models based on CNN Abbreviations: IVC, internal validation cohort; EVC, external validation cohort; Acc, accuracy; Sen, sensitivity; Spe, specificity; PPV, positive prediction value; NPV, negative prediction value Table 4DeLong_testCombined_ViT Vs CE_T2FLAIR_ViTCombined_ViT Vs CE_T1WICombined_ViT Vs Combined_ShuffleNetCombined_ViT Vs Combined_Resnet50Combined_ViT Vs Combined_DesNet121cohort P value < 0.001 < 0.001 0.029 0.008 0.008TrainingP value 0.027 0.036 0.043 0.046 0.048 IVCP value 0.003 0.005 0.015 0.018 0.024 EVC Combined_DesNet121 Vs CE_T2FLAIR_ViT Combined_DesNet121 Vs CE_T1WI Combined_DesNet121 Vs Combined_ShuffleNet Combined_DesNet121 Vs Combined_Resnet50 cohort P value 0.006 0.017 0.423 0.784 TrainingP value 0.085 0.115 0.799 0.891 IVCP value 0.044 0.218 0.686 0.847 EVC Combined_Resnet50 Vs CE_T2FLAIR_ViT Combined_Resnet50 Vs CE_T1WI Combined_Resnet50 Vs Combined_ShuffleNet cohort P value 0.015 0.035 0.622 TrainingP value 0.086 0.126 0.875 IVCP value 0.117 0.201 0.839 EVC Combined_ShuffleNet Vs CE_T2FLAIR_ViT Combined_ShuffleNet Vs CE_T1WI cohort P value 0.039 0.107 TrainingP value 0.086 0.146 IVCP value 0.119 0.407 EVC CE_T1WI Vs CE_T2FLAIR_ViT cohort P value 0.586 TrainingP value 0.889 IVCP value 0.508 EVCAbbreviations: IVC, internal validation cohort; EVC, external validation cohort DeLong_test Abbreviations: IVC, internal validation cohort; EVC, external validation cohort DCA of different models. (a) prediction results in IVC, (b) prediction results in EVC The calibration curve of different models. (a) prediction results in IVC, (b) prediction results in EVC Discussion Accurate assessment of glioma proliferative activity is essential for effective patient management. Ki-67, a well-established biomarker of cell proliferation, is widely used in the grading and prognostic evaluation of central nervous system tumors [21]. For patients who cannot undergo surgery, preoperative biopsy remains the gold standard for determining Ki-67 expression levels in gliomas. However, its invasive nature, potential complications, and high cost restrict its broad clinical application. Therefore, preoperative imaging-based prediction of Ki-67 expression carries substantial clinical importance. With the advancement of medical image processing methods and the research in artificial intelligence, the study of DL methods in disease detection and identification has continuously deepened, and the methods have been widely applied in tumor diagnosis and prognosis assessment [22–24]. At the same time, they have also provided new approaches for predicting the grade and molecular classification of gliomas before surgery [25–27]. Among the methods, CNNs are typical DL artificial neural networks capable of directly capturing and extracting features from MRI images that are unobservable to the human eye. In medical image analysis, CNNs inherently possess translation invariance, which endows them with natural robustness against geometric transformations of input data—including rotation, scaling, and translation [28]. In addition to classical CNN methods, new DL models developed in recent years have revitalized medical image analysis, such as the ViT. ViT is a novel attention-based model built on the self-attention mechanism. On the one hand, self-attention mechanism can capture complex image features that CNNs often overlook; on the other hand, it can automatically learn feature relationships across different image regions, thereby enhancing model generalizability. This has made it a core model in the field of natural language processing (NLP) [29]. While CNNs, composed of convolutional layers, pooling layers, and fully connected layers, exhibit strong local feature representation, they struggle to capture long-range spatial relationships and global contextual information [10]. Due to the unique advantages of ViT over CNNs, ViT has increasingly been applied in the medical field in recent years, including medical image segmentation, molecular classification prediction, and tumor prognosis, where it has achieved substantial progress. However, to date, no study has constructed a transformer-based DL model to predict the Ki-67 expression level in gliomas. Therefore, the findings of our study further demonstrate the potential of the ViT model in predicting the molecular subtypes of gliomas. This study first constructed two single-modal (CE-T1WI and CE-T2FLAIR) DL models based on a transformer architecture to predict the Ki-67 expression level of gliomas. Subsequently, a combined model was developed by integrating CE-T1WI and CE-T2FLAIR. The results demonstrated that all three transformers-based model were effective in predicting the Ki-67 expression level of gliomas, with the combined model achieving the best predictive performance, as corroborated by the DeLong test results presented in Table 4. Additionally, we observed that the CE-T1WI_ViT model showed a slightly higher AUC in the EVC than in the IVC (0.859 vs. 0.837). This small numerical difference may be related to cohort-level variability, including differences in sample composition, imaging characteristics, lesion heterogeneity, and Ki-67 class distribution. Therefore, the performance of this model should be interpreted as generally stable across validation cohorts, rather than indicating a meaningful deterioration or improvement. Yunyang Zhu et al. [30] enrolled 387 patients with primary glioma from three centers, extracted imaging features via radiomics, deep learning and habitat analysis, and constructed various models for Ki-67 prediction. They found that radiomics(AUC:0.931), habitat-based(AUC:0.855) models, and various combined models all enabled satisfactory prediction of Ki-67 expression in gliomas. Among these models, the one based on the LightGBM modeling algorithm combined with ResNet50 features from habitat subregions achieved the best performance in predicting Ki-67 expression levels in gliomas, outperforming other single or combined models.Notably, the transformer-based model in our study achieved a slightly higher AUC than the model developed by Yunyang Zhu et al. [30], especially compared with the habitat-based model. However, there are differences in the region delineation between our study and theirs: their habitat model incorporated regions from the “hyperintensity area” and “non-hyperintensity area” of CE-T1WI, as well as the “tumor area” and “edema area” of T2WI sequences. In contrast, during our region delineation, we only included the “hyperintensity area” and “non-hyperintensity area” on CE-T1WI and CE-T2FLAIR, without considering the inclusion of peritumoral edema. This is one of the limitations of our study, because for gliomas, particularly high-grade ones with IDH-wildtype status, the infiltrative margin and peritumoral zone carry important information about Ki-67 proliferative activity [31], which may have affected the further improvement of our model’s prediction performance. In addition to Ki-67 prediction, the ViT model also showed strong performance in predicting other molecular subtypes of gliomas. Yu X et al. [32] collected 356 preoperative T1WI and CE-T1WI MR images of glioblastoma patients and developed DL model for tumor and peritumoral edema using the transformers algorithm to predict O6-methylguanine-DNA methyltransferase (MGMT) promoter methylation status. Their results indicated that both the single ViT model (for tumor or peritumoral edema) and the combined model were capable of predicting MGMT promoter methylation status, with the combined model demonstrating superior predictive performance (AUC: 0.923, 95% CI: 0.890–0.948). A recent study [33] further demonstrated the potential of the ViT model in predicting the Ki-67 expression level of meningiomas. Chaoyue Chen et al. [33] collected multimodal MR Image data from 1239 meningioma patients at two medical institutions between January 2017 and December 2023. By integrating DL features, radiological features, and Radiomics features of MR images through three transformers modules, they constructed a cascaded multimodal model to predict the Ki-67 index of meningioma patients. The model achieved an AUC of 0. 808 in the EVC. Importantly, this model enables the identification of asymptomatic meningioma patients who require early intervention, providing significant clinical value for the personalized treatment of meningioma patients. Finally, to compare the predictive performance of the ViT model and the CNN model in estimating the Ki-67 expression level of glioma, this study trained three CNN models (ShuffleNet, ResNet50, and DenseNet121) using CE-T1WI and CE-T2FLAIR data from 421 patients, and compared them with our trained Combined_ViT model. The the DeLong test results showed that the Combined_ViT model outperformed all three CNN models, demonstrating the advantages of the ViT model over CNN in global receptive field coverage and long-distance dependency modeling. Byeon Y et al. [34] collected multimodal MR images (T1WI, T2WI, T2FLAIR, CE-T1WI) of 1, 053 patients with adult diffuse glioma from three centers and combined the multimodal slices into a four-channel input image. Using CNN and ViT, they constructed DL model for multiple tasks (IDH mutation, 1p/19q co-deletion, tumor grade). Their results showed that the AUC values of the ViT model across all tasks were higher than those of the CNN model, confirming the superiority of ViT over CNN in multi-task predictive. However, their model were 2D rather than 3D, whereas our study developed a 3D DL model. Furthermore, the 3D DL model in our study achieved slightly higher AUC than theirs in predicting the molecular classification of glioma, likely because the 3D model preserved complete information, including tumor shape and spatial relationships, thereby improving predictive accuracy. Huang J et al. [35] collected laboratory parameters from 506 glioma patients within one week before surgery across two medical institutions and used extreme gradient boosting (XGBoost), support vector machine (SVM), and least absolute shrinkage and selection operator (LASSO) to screen parameters associated with Ki-67 levels. Ultimately, 15 laboratory parameters were selected. They then used 10 machine learning classifiers to construct model for predicting Ki-67 expression levels in glioma. The logistic regression (LR) model achieved the best performance, with an AUC of 0.800 in the IVC and 0.757 in the EVC. Their study demonstrated that machine learning models constructed based on laboratory parameters can also achieve good performance in predicting the Ki-67 expression level of gliomas besides DL models. Furthermore, we constructed a ViT model based on CE-T2FLAIR. CE-T2FLAIR images are strongly T2-weighted, with subtle T1 effects resulting from prolonged inversion times, which also allow for gadolinium-related enhancement. This Image modality captures both T2 signals and contrast-contrastenhanced lesions visible on T1-weighted images, thereby providing a comprehensive representation of tumor and surrounding tissue characteristics. Numerous studies have reported the diagnostic advantages of CE-T2FLAIR in brain lesion assessment [36]. At the same time, our research also demonstrates the benefits of CE-T2FLAIR in the development of DL model. To the best of our knowledge, this is the first DL study based on CE-T2FLAIR to date. Previous studies [34] have constructed predictive DL model mainly using conventional sequences such as T2WI, T2FLAIR, and CE-T1WI. Although this study achieved satisfactory results, it still has several limitations. First, downsampling Center 1 data from 1 mm to 5 mm slice thickness to match Center 2 may introduce potential information loss, although our small-sample quantitative analysis results showed extremely high consistency between the segmentation results of the original 1 mm images and the resampled 5 mm images. Second, this study established the DL model using only CE-T1WI and CE-T2FLAIR, without incorporating functional sequences such as diffusion weighted imaging(DWI), susceptibility-weighted imaging (SWI) and diffusion kurtosis imaging (DKI). We plan to include these functional sequences in our future studies—we will collaborate with more clinical centers to standardize the acquisition of functional MRI sequences, ensure data consistency, and further improve the biological explainability and clinical applicability of our Ki-67 prediction model. Finally, given that the infiltrative margin and peritumoral zone carry information about Ki-67 proliferative activity in gliomas, we plan to incorporate the peritumoral edema region into our future research to further improve the prediction performance of our model. In conclusion, both transformer-based and CNN-based deep learning model can predict the Ki-67 expression level of grade 2–4 gliomas prior to surgery, with the ViT model demonstrating superior predictive performance compared to the CNN model. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 2 Acknowledgements Not applicable. Author contributions Liu Liu completed experimental design, data collection, image segmentation and original manuscript. Fangfang Hu assisted in data collection and image segmentation. Minquan Lai provided technical support in the field of artificial intelligence. Xianjing Zhao and Pinggui Lei reviewed and edited the manuscript. Zhenwei Yao and Bing Fan helped to design the work, revised the final version of the manuscript and provided funding support. All authors read and approved the final manuscript. Funding This study was supported by the National Natural Science Foundation of China (Grant No.82272061) and Jiangxi Provincial Natural Science Foundation(Grant No. 20202BABL206112). Data availability All data supporting the conclusions of this article will be made available by the authors. Further enquiries can be directed to the corresponding author. Declarations Ethics approval and consent to participate This study was approved by the Medical Ethics Committee of Jiangxi Provincial People’s Hospital and the requirement for informed consent was waived. The ethical number was 2024 (68). Our research was conducted in accordance with the Declaration of Helsinki. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Zhenwei Yao, Email: zwyao@fudan.edu.cn. Bing Fan, Email: 26171381@qq.com. References Associated Data Supplementary Materials Supplementary Material 1 Supplementary Material 2 Data Availability Statement All data supporting the conclusions of this article will be made available by the authors. Further enquiries can be directed to the corresponding author.