Intratumor habitat aware deep learning framework for preoperative lymph node metastasis prediction in laryngeal carcinoma Authors: Wencan Wang, Lingyu Zhang, Junhong Zhang, Xinyu Li, Wenfeng Wei, Li Liang, Shiwen Li, Haomiao Ma, Xuhui Zhao, Qian Zhang, Huina Guo, Yanhong Hao, Hua Zhang, Feng Gao, Jianxin Zhang, Lingjie Wang Published: 2026 ABSTRACT AbstractObjectivesTo develop an intratumor heterogeneity (ITH)-aware deep learning (DL) framework for preoperative prediction of lymph node metastasis (LNM) in laryngeal squamous cell carcinoma (LSCC).MethodsThis multi-center retrospective study, which included 381 patients, proposed a ITH-aware DL framework integrated with adaptive habitat mapping to automatically decode spatial ITH from primary tumor contrast-enhanced CT images.The predictive performance of the ITH-aware DL model was compared against conventional radiomics and 3D DL model. A stacking ensemble model combining these three methods was also constructed. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), calibration curves, and decision curve analysis (DCA). The prognostic value for overall survival (OS) stratification was evaluated.ResultsThe ITH-aware DL model achieved superior preoperative LNM prediction in external testing (AUC: 0.745–0.766), outperforming both radiomics (AUC: 0.682–0.694) and 3D DL models (AUC: 0.718–0.732). The stacking ensemble integrating these models attained the highest diagnostic accuracy (AUC: 0.786–0.804). Critically, both the ITH-aware DL and stacking models effectively stratified patients into distinct risk groups with significantly different OS outcomes.ConclusionThe proposed ITH-aware DL framework provides a robust, automated tool for preoperative LNM prediction in LSCC. By enabling accurate risk stratification, it holds significant potential to inform personalized surgical planning and improve treatment strategies.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s40644-026-01096-8.Keywords: Laryngeal carcinoma, Lymph node metastasis, Intratumoural heterogeneity, Deep learning, Survival Introduction Laryngeal carcinoma constitutes one of the most prevalent respiratory malignancies worldwide, with a notably increasing incidence among males, where it ranks as the 11th most common cancer globally [1]. Over 85% of laryngeal carcinoma cases are identified as laryngeal squamous cell carcinomas (LSCC), which are characterized by their aggressive biological behavior, notably early lymphatic dissemination, leading to poor clinical outcomes [2]. Recent studies have highlighted that the therapeutic response of metastatic lymph nodes (LNM) significantly influences overall treatment efficacy in LSCC management, more so than the response of the primary tumor itself. Insufficient control or recurrence of LNM post-treatment often results in regional treatment failure, elevated rates of distant metastasis, and diminished patient survival [3, 4]. Therefore, precise preoperative prediction of LNM is crucial for accurate TNM staging, personalized treatment planning, and survival optimization [5–8]. Traditional approaches to LNM assessment face significant limitations. Although histopathological biopsy is considered the diagnostic gold standard, its invasive nature and associated procedural risks restrict its routine application. Moreover, sampling limitations in heterogeneous LNM can lead to false-negative results [9, 10]. Conventional imaging modalities—such as CT, MRI, and PET—provide essential morphological and metabolic information [11, 12]; however, their sensitivity in detecting subcentimeter LNM remains suboptimal, and their reliance on size and morphological criteria can also yield false-positive diagnoses by mistaking enlarged inflammatory nodes for metastases. Studies have reported that conventional MRI detects only approximately 35% of subcentimeter LNM [9, 13]. Current radiological evaluations still rely heavily on subjective visual interpretation by radiologists, an approach prone to inter-observer variability and diagnostic inconsistency [14]. Accumulating evidence underscores the critical role of spatial heterogeneity within the primary tumor as a key determinant of metastatic propensity in LSCC [15, 16]. This heterogeneity, manifested as spatially organized variations in cellular phenotypes, vascular architecture, and molecular profiles within the tumor ecosystem, fundamentally governs lymphatic dissemination and disease progression. Consequently, comprehensive mapping of intratumoral heterogeneity (ITH) is essential not only for reliable prediction of metastasis but also for accurate prognostic stratification.Advanced computational analysis of primary lesions offers considerable potential to uncover subtle yet prognostically decisive imaging signatures. In this context, radiomics has emerged as a valuable approach for non-invasive tumor evaluation, leveraging high-throughput extraction of subvisual textural features [17, 18]. This methodology has exhibited promising predictive performance across various clinical scenarios, including predicting disease progression during chemoradiotherapy for oropharyngeal carcinoma and identifying pathological complete responses following neoadjuvant immunochemotherapy in non-small cell lung cancer [11, 19]. Nevertheless, conventional radiomics approaches often fall short in comprehensively characterizing spatially organized ITH, primarily due to their tendency to treat tumors as homogeneous entities or their reliance on manually segmented subregions of uncertain biological relevance. Deep learning (DL) methodologies, in contrast, offer distinct advantages in decoding complex high-dimensional imaging patterns. DL has demonstrated extensive utility across multiple facets of disease management, including diagnostic applications such as fracture risk prediction in osteoporosis [20], prognostic tasks such as survival stratification in colorectal cancer [21], and therapeutic response assessment such as predicting major pathological response to neoadjuvant chemoimmunotherapy in non-small cell lung cancer [22]. These architectures minimize manual processing requirements while supporting integrated, end-to-end image analysis. However, despite these capabilities, to our knowledge, no prior studies have successfully implemented a DL framework for decoding spatial ITH in the specific context of LNM prediction for LSCC. In this study, we present a novel convolutional neural network (CNN) framework, termed the Intratumor Heterogeneity-Aware Deep Learning (ITH-aware DL) framework, which incorporates adaptive habitat mapping to automatically decode ITH from primary tumor regions. This DL approach enhances the preoperative prediction of LNM in LSCC.Given that the prognostic significance of LNM in LSCC has been confirmed, we further hypothesized that the LNM probability predicted by the ITH-aware DL model could effectively stratify patients according to overall survival (OS) risk, thereby providing a clinical tool to link preoperative imaging with long outcomes. Materials and methods Ethics statement This retrospective study was conducted in strict accordance with the Helsinki Declaration. It has received ethical approval from the Institutional Review Board of the First Hospital of Shanxi Medical University(Approval No.KYYJ-2025-59). Since the dataset has been de-identified and complies with the regulatory guidelines for the use of secondary data, no written informed consent form needs to be signed. Patients This retrospective study collected data on patients with LSCC from 2016 to 2024.The inclusion criteria were as follows: (1) histopathologically confirmed LSCC and has undergone cervical lymph node dissection.; (2) neck contrast-enhanced computed tomography (CECT) performed within two weeks prior to surgery; (3) the initial treatment was surgery without prior neoadjuvant radiotherapy; and (4) absence of distant metastasis or multiple primary malignancies. The exclusion criteria included: (1) non-squamous cell carcinoma pathology; (2) presence of metal or respiratory motion artifacts in CT images that impaired lesion segmentation; and (3) incomplete clinical baseline data, including age, gender, and TNM stage.The primary outcome of this study is the presence of cervical LNM, which is defined as the detection of metastatic carcinoma in one or more surgically resected cervical lymph nodes confirmed by histopathology.This histopathological diagnosis serves as the gold standard reference.In this study, 219 patients (Center 1, n = 219) from the First Affiliated Hospital of Shanxi Medical University, enrolled between June 2018 and April 2024, were randomly assigned to a training set and an internal test set at a ratio of 7:3. Additionally, 162 patients from the Shanxi Provincial Cancer Hospital were allocated into two external test sets based on the time of data collection: patients enrolled from January 2016 to December 2018 (Center 2−1, n = 88) formed External Test Set 1, and those from January 2019 to December 2020 (Center 2–2, n = 74) constituted External Test Set 2, which was also used for survival analysis. The patient inclusion process is seen in Fig. 1. Flowchart of patient recruitment CT image processing and segmentation The research flowchart is presented in Fig. 2. All patients underwent CECT scans prior to surgery. Detailed information regarding image acquisition parameters and processing procedures is provided in Supplementary Appendix S1 and Table S1. Two head and neck radiologists, each with over 10 years of experience, independently performed image segmentation while being blinded to all clinical data. The regions of interest (ROI) were manually delineated layer-by-layer using the open-source software ITK-SNAP (v3.8.0). To evaluate the reliability and consistency of ROI delineation, the ROI of 50 randomly selected patients were re-labeled after a one-month interval, and intraclass correlation coefficients (ICC > 0.85) were calculated to quantify inter-and intra-observer agreement. Flowchart of the study depicting the sequential workflow: image segmentation, construction of ITH-aware deep learning model, 3D deep learning model, radiomics feature extraction, assessment of model performance, and survival analysis ITH-aware DL model development and feature extraction Adaptive habitat mapping Habitat mapping construction is performed using the supervoxel clustering algorithm.The optimal number of clusters for each tumor was adaptively determined by maximizing the silhouette coefficient within the range of 5 to 50, eliminating the need for manual pre-definition. For details, see Supplementary Appendix S1. ITH-aware DL image integration For each patient, through adaptive clustering technology, 3D tumor region segmentation was carried out throughout the entire tumor volume, the maximum cross-sectional slices were extracted from the original CT images and the 3D clustering map respectively. These two slices were integrated into a dual-channel input and were input into the CNN. The minimum rectangle encompassing all tumor regions across the two slices was defined as the ROI for the ITH-aware DL input. ITH-aware DL model training To enhance the generalization ability and performance of the model, multiple data augmentation techniques were implemented, including random cropping, horizontal flipping, vertical flipping, rotation, and scaling. Each channel was standardized to a spatial resolution of 224 × 224 pixels. The ResNet18 architecture was pre-trained on the ImageNet dataset and subsequently fine-tuned on our training set using transfer learning. To adapt ResNet18 for the LNM prediction task, a global fine-tuning strategy was applied to optimize the network parameters. Upon completion of model training, 512 DL features were extracted from the average pooling layer of the ResNet18 architecture for each image block.Since the habitat map is integrated into the ResNet18 network as a second input channel, variations in the number of habitats do not affect the dimension of the extracted feature vectors. Additionally, the Grad-CAM technique was employed to visualize the model’s decision-making process, highlighting the key tumor regions that were identified in an interpretable manner. These visualization results are intended to support clinicians in prognosis assessment and clinical decision-making. The DL model Hyperparameter Configuration are detailed in Supplementary Appendix S1. 3D DL model development and feature extraction Initially, a 3D image encompassing the entire tumor volume was selected as the model input. Subsequently, the smallest cubic region containing the tumor was extracted. Finally, this cubic region was resized to a standardized dimension of 48 × 48 × 16 pixels to serve as the final input for the model. During the training phase, data augmentation techniques were applied to enhance dataset diversity and mitigate the risk of overfitting. These techniques included rotational, translational, and scaling transformations, which effectively increased the number of training samples. Similarly, the ResNet18 architecture was implemented and trained for 200 epochs with a learning rate of 0.01. The trained 3D CNN was capable of predicting the LNM status of patients. To extract the 3D DL features for each patient, 512 DL features were derived from the average pooling layer of the 3D ResNet18 architecture. Radiomics feature extraction A total of 1834 radiomics features were extracted from each patient’s tumor ROI using Pyradiomics (v3.1.0). These features comprised 14 shape-based features, 360 first-order statistical features, 280 Gy-level dependence matrix (GLDM) features, 320 Gy-level size zone matrix (GLSZM) features, 100 neighborhood gray-level difference matrix (NGTDM) features, 320 Gy-level run-length matrix (GLRLM) features, and 440 Gy-level co-occurrence matrix (GLCM) features. The feature distribution map is in Supplementary Figure S1. Feature selection and model construction Feature selection was independently performed on the radiomics features, ITH-aware DL features, and 3D DL features. Feature standardization was conducted using the Z-score method. The feature selection process consisted of three sequential steps: (1) Features with high repeatability (ICC > 0.85) were selected based on ICC analysis. (2) To reduce redundancy and multicollinearity among features, Pearson correlation analysis was further applied. (3) Within a 5-fold cross-validation framework, Lasso regression was employed to identify the most discriminative feature subset, aiming to optimize feature selection and enhance the model’s discriminative capability. The regularization parameter (λ) was tuned to maximize inter-class separation.To address the issue of sample imbalance caused by the small sample size of the LNM-positive group (n = 92), the SMOTE algorithm was applied in the model training stage after data partitioning. For each feature set, six classifiers (LR, SVM, Random Forest, XGBoost, Extra Trees, and AdaBoost) were evaluated on the training set via 5-fold cross-validation. Among all candidate classifiers, LR exhibits a smaller gap between training performance and validation performance, indicating excellent generalization stability. Furthermore, the predicted probability output generated by LR is well-calibrated, which is critical for the subsequent stacking ensemble and patient risk stratification. Construction and assessment of the stacking model An ensemble learning strategy was employed to integrate the output scores of each unimodal model, which were then used as input features for training the same classifier utilized in the individual models. The LR classifier, identified as the best-performing model, was selected to construct the stacking ensemble model. The predictive performance of the model was visually evaluated using the receiver operating characteristic (ROC) curve, calibration curve, and decision curve analysis (DCA). Diagnostic performance metrics of the three unimodal models and the stacking model were calculated, including the area under the curve (AUC), 95% confidence interval (95% CI), specificity, sensitivity, accuracy, positive predictive value (PPV), and negative predictive value (NPV). Survival analysis Since external test set 2 contained complete follow-up data, we evaluated the prognostic value of each model using this dataset. The primary endpoint of the follow-up was OS, defined as the time interval from the initiation of treatment to either death or the last recorded follow-up. Follow-up data were obtained from both inpatient and outpatient medical records. According to the follow-up protocol, patients were scheduled for follow-up visits every 3–6 months during the first two years after surgery, and annually thereafter. The follow-up period was censored as of December 31, 2023. Using the maximum cutoff values derived from the Youden index of each model, patients in external test set 2 were categorized into high-risk and low-risk groups. Kaplan-Meier (K-M) survival curves were generated to assess prognostic differences, and the log-rank test was applied to determine statistically significant differences between survival curves. Statistical analysis The normality of clinical features was confirmed using the Shapiro-Wilk test. Continuous variables were analysed via distribution-agnostic hypothesis testing (t-test or Mann–Whitney U test), while categorical variables underwent chi-square (χ²) evaluation. All analyses were performed on the OnekeyAI platform (v4.9.1, Python 3.7.12) using an integrated technical stack: Statsmodels (v0.13.2) for statistical modelling, Pyradiomics (v3.1.0) for radiomics feature extraction, Scikit-learn (v1.0.2) for machine learning implementation, and PyTorch (v1.11.0) with CUDA (v11.3.1) and cuDNN (v8.2.1) optimisations for DL workflows. This toolchain ensures reproducible, high-performance computational analysis.K-M survival analysis was performed using SPSS (Version 27.0). All statistical tests were two-tailed, with a P-value < 0.05 considered statistically significant. Results Patient characteristics analysis The clinical characteristics of the patients are shown in Table 1. A total of 381 patients with LSCC were included in this study, including 153 cases in the training set, 66 cases in the internal validation set, 88 cases in external test set 1, and 74 cases in external test set 2. Among the patient population, there were 289 cases (75.8%) with negative lymph nodes and 92 cases (24.2%) with positive lymph nodes. Table 1Characteristics of patients in the datasetsVariableTraining set(n = 153)Internal validation set (n = 66)External test set 1(n = 88)External test set 2(n = 74)p valueLNM status0.56 LNM-negative120(78.43)47(71.21)64(72.73)58(78.38) LNM-positive33(28.79)19(28.79)24(27.27)16(21.62)Age62.87 ± 8.7662.88 ± 8.4661.18 ± 8.0661.85 ± 7.090.41Gender0.91 Male147(96.08)62(93.94)84(95.45)71(95.95) Female6(3.92)4(6.06)4(4.55)3(4.05)Smoking<0.001 Yes118(77.12)54(81.82)71(80.68)35(47.30) No35(22.88)12(18.18)17(19.32)39(52.70)Alcohol drinking0.001 Yes65(42.48)34(51.52)30(65.91)16(21.62) No88(57.52)32(48.48)58(65.91)58(78.38)T stage0.02 T159(39.56)22(33.33)32(36.36)27(36.49) T234(22.22)17(25.76)27(30.68)24(32.43) T348(31.37)25(37.88)14(15.91)16(21.62) T412(7.84)2(3.03)15(17.05)7(9.49)Clinical stage0.03 I53(34.64)22(33.33)27(30.68)21(28.38) II13(8.50)9(13.64)20(22.73)13(17.57) III44(28.76)27(40.91)20(22.73)21(28.38) IV43(28.10)8(12.12)21(23.86)19(25.68)Surgical approach< 0.001 Partial117(76.47)52(78.79)63(71.59)36(48.65) Total36(23.53)14(21.21)25(28.41)38(51.35)Chemoradiotherapy0.11 Yes113(73.86)47(71.21)52(59.09)49(66.22) No40(26.14)19(28.79)36(40.91)25(33.78) Characteristics of patients in the datasets Feature extraction and selection A total of 1834 radiomics features, 512 ITH-aware DL features and 512 3D DL features were extracted. After LASSO feature selection, 16 ITH-aware DL features, 4 3D DL features and 12 radiomics features were respectively screened out. Among the various selected features, the number of ITH-aware DL features was the largest and the weights of each feature were generally higher. The visualization results of the weights of various features are shown in Supplementary Figure S2-S4. t-SNE dimension reduction was used for visualization of these features. Development and performance of models For these three individual feature sets, six classifiers were evaluated. LR consistently demonstrated the best predictive performance across all feature sets. The outcomes of the six machine learning algorithms employed by each model are presented in the Supplementary Table S2-S4. Based on the output scores of these three individual models, LR also exhibited superior predictive accuracy (Table S5). The ROC curves of the six classifiers in each model are presented in Supplementary Figure S5-S8. Among the three distinct modalities, the ITH-aware DL model achieved the higher AUC values in all study sets (training set: 0.910, internal validation set: 0.720, external test set 1: 0.745, external test set 2: 0.766). The stacking model integrating these three distinct modalities achieved the highest AUC values in the training set and external test set 1. The corresponding AUC values are as follows (training set: 0.935, internal validation set: 0.767, external test set 1: 0.804, and external test set 2: 0.786). Figure 3 shows the ROC curves of different models in the four sets. The performance metrics of the four models across various research sets are summarized in Table 2.The DCA curves in the research set indicate that the stacking model offers greater net benefit. Additionally, the calibration curves of the three distinct modalities show good consistency over a wider range of probabilities compared to the stacking model. The DCA curves and calibration curves are shown in Figs. 4 and 5. The pixel weight distribution of the Grad-CAM ITH-aware DL model (Fig. 7) provides interpretable insights for preoperative prediction of LNM in laryngeal carcinoma. Receiver operating characteristic curves of different models in the training set (A), internal validation set (B), external test set 1 (C), and external test set 2 (D) Table 2The performance comparison of different modelsModelSetAUC (95%CI)AccuracySensitivitySpecificityPPVNPVITH-aware DLTraining0.910 [0.856–0.965]0.8760.7580.9080.6940.932Validation0.720 [0.594–0.847]0.6210.9470.4890.4290.958Test10.745 [0.624–0.867]0.7500.6250.7970.5360.850Test20.766 [0.609–0.923]0.8510.5620.9310.6920.8853D DLTraining0.828 [0.759–0.897]0.7120.8480.6750.4180.942Validation0.844 [0.748–0.941]0.7730.8420.7450.5710.921Test10.732 [0.621–0.844]0.6930.7500.6720.4620.878Test20.718 [0.587–0.849]0.5540.9370.4480.3190.963RadiomicsTraining0.899 [0.847–0.952]0.7520.9390.7000.4630.977Validation0.783 [0.665–0.901]0.7120.8950.6380.5000.937Test10.694 [0.580–0.808]0.6700.7920.6250.4420.889Test20.682 [0.548–0.815]0.5680.8750.4830.3180.933StackingTraining0.935 [0.885–0.984]0.8820.8480.8920.6830.955Validation0.767 [0.642–0.892]0.7420.6840.7660.5420.857Test10.804 [0.710–0.898]0.7160.7920.6870.4870.898Test20.786 [0.662–0.909]0.7430.7500.7410.4440.915 The performance comparison of different models Calibration curves for the models in the training set (A), internal validation set (B), external test set 1 (C), and external test set 2 (D) DCA curves for the models in the training set (A), internal validation set (B), external test set 1 (C), and external test set 2 (D) Prognostic risk stratification of models Based on the output probability of the four models on the external test set 2, we uniformly divided the study set into high-risk and low-risk groups. The K-M survival curves indicated that in the external test set 2, the ITH-aware DL models (log-rank test P = 0.0002) and the stacking model (log-rank test P = 0.0006) could significantly stratify patients into different groups with distinct OS outcomes. The 3D DL model (log-rank test P = 0.0023) and the radiomics model (log-rank test P = 0.0187) also achieved statistically significant stratification. These results indicate that all four models have prognostic value, among which the ITH-aware DL model and the stacking model demonstrated the stronger survival discrimination ability. Figure 6 shows the K-M curves of the four models. Subgroup analysis of survival stratification To evaluate the robustness of prognostic stratification across different clinical subgroups, we performed K-M survival subgroup analysis on the external test set 2 stratified by age (≤ 60 years and > 60 years) and T stage (T1–T2 and T3–T4). Both the ITH-aware DL model and the stacking model consistently stratified patients into groups with significantly distinct OS across all subgroups. Survival curves for each subgroup are presented in Supplementary Figure S12 and S13. These findings confirm that the prognostic value of the ITH-aware DL model and the stacking model is not restricted to specific clinical subgroups, but demonstrates broad applicability. Discussion Accurate preoperative prediction of LNM is crucial for optimizing treatment strategies in LSCC. In this study, we developed a DL framework that integrates supervoxel segmentation, adaptive dynamic clustering, and a CNN to decode ITH from preoperative CECT images for LNM prediction in laryngeal carcinoma. This framework automatically characterized ITH and, based on the resulting habitat maps, demonstrated robust predictive performance across both training and external test sets, achieving a maximum AUC of 0.766, surpassing conventional radiomics and 3D DL models. Furthermore, by integrating predictions from radiomics, 3D DL, and this ITH-aware DL framework via a stacking ensemble strategy, we further enhanced predictive accuracy. Importantly, this ITH-informed risk stratification effectively categorized patients into low-risk and high-risk groups with significantly distinct OS. ITH analysis has become a crucial factor in radiology for preoperative risk assessment and treatment response prediction. Numerous studies have demonstrated its clinical value across various cancers. For instance, Wang et al. employed unsupervised clustering to partition bladder tumors into three subregions for risk stratification (AUC = 0.832) [23]; Xie et al. utilized subregional radiomics to predict survival in esophageal cancer patients (AUC = 0.805) [24]; and Wu et al. delineated tumor subregions via PET and CECT clustering to predict progression in oropharyngeal carcinoma [11]. Moreover, multiple studies have shown that ITH is a key driver of LNM in LSCC, and its spatial distribution patterns within the primary tumor can significantly influence the propensity for lymphatic dissemination [3, 6–8, 25]. In current clinical practice, the diagnosis of cervical LNM in LSCC relies primarily on the subjective visual interpretation of radiologists, an approach inherently associated with inter-observer variability and experience-based discrepancies. Therefore, accurate quantification and characterization of ITH hold significant value for predicting LNM and overcoming the limitations of conventional diagnostic approaches. Despite the promising results of prior image-based ITH studies, these methods methods exhibit notable limitations in image segmentation strategies: (1) Voxel-based radiomics feature extraction is computationally intensive and difficult to implement in routine clinical practice [26, 27]; (2) Reliance on manually preset cluster numbers fails to adaptively capture the heterogeneity of ITH across different tumors, potentially leading to biologically irrelevant clustering [28, 29]. To address these issues, we innovatively introduced ITH-aware DL framework that not only automatically identifies tumor subregions but also autonomously determines the optimal number of partitions based on the heterogeneity level of each tumor, eliminating the need for predefined cluster numbers. Furthermore, this ITH-aware DL framework demonstrated favorable performance in predicting LNM in this study.To elucidate the decision-making process of the ITH-aware DL model, this study generated Grad-CAM (Fig. 6). Observations reveal that there are significant differences in attention patterns between metastatic and non-metastatic cases. In N+ cases, the ITH-aware DL model focuses on the central subregion of the tumor; at the clinical level, these high-attention local regions typically correspond to dense solid tumor cores or regions with high heterogeneity. In contrast, N0 cases exhibit a scattered and unfocused attention pattern with no distinct invasive core. Based on the maximum cutoff values derived from the Youden index of each model, patients in the external test set 2 were consistently stratified into high-risk and low-risk groups (the blue curve represents the low-risk group, the orange curve denotes the high-risk group). Panels A-D illustrate the Kaplan-Meier curves for different models: the radiomics model (A), the 3D deep learning model (B), the ITH-aware deep learning model (C), and the stacking model (D) Grad-CAM visualization for model interpretation. a and b show the activation maps generated by the ITH-aware DL model for patients with positive lymph nodes in laryngeal cancer, while c and d present the feature activation maps of the same ITH-aware DL model for patients with negative lymph nodes in laryngeal carcinoma Conventional radiomics and 3D DL techniques, which evaluate the entire tumor volume, have been extensively applied in oncology for diagnosis, treatment assessment, and prognosis prediction [17, 21, 22, 30–32], including LNM prediction in LSCC [33]. In our study, both approaches demonstrated acceptable predictive performance, but were outperformed by the ITH-aware DL model. Notably, among the 12 radiomics features retained in the final model—six texture and six first-order features—a majority were associated with heterogeneity, further supporting the role of ITH in LNM. Although ITH primarily captures local tumor attributes, previous research, such as in breast cancer, has demonstrated that combining ITH with global tumor representations enhances predictive power [34]. Therefore, we combined radiomics, 3D DL, and ITH-aware DL models into a stacking ensemble framework, which achieved superior performance, likely due to its ability to capture both localized subregional heterogeneity and comprehensive tumor-wide information. Given the established correlation between LNM and poor prognosis in LSCC, our DL model offers a clinically aligned approach to preoperative risk stratification: high-risk patients may benefit from endoscopic resection or elective neck irradiation, whereas low-risk patients could avoid unnecessary surgical intervention [33]. Critically, patients identified as high-risk by both the ITH-aware DL model and the stacking ensemble model showed significantly shorter OS (log-rank P < 0.005). The subgroup analysis further demonstrated that the prognostic value of both the ITH-aware DL and stacking models was robust across age and T stage subgroups.This survival differentiation, derived directly from model predictions, confirms the prognostic importance of LNM status and underscores the potential of our ITH-aware DL framework to enhance preoperative evaluation and facilitate personalized treatment strategies. Several limitations of this study should be acknowledged. First, the retrospective design may introduce selection bias; therefore, future prospective studies are necessary to validate the model’s performance. Second, manual segmentation of the ROI represents a potential source of variability and inter-observer subjectivity. However, high ICC values (ICC > 0.85) confirmed strong reproducibility, and the models consistent performance across two independent external validation sets suggests that manual segmentation had a limited impact on our conclusions. The future adoption of automated segmentation methods will further enhance objectivity and reproducibility. Furthermore, we plan to expand the cohort size and integrate multimodal data—including MRI, ultrasound, genomic, and histopathologic features—to more comprehensively characterize the relationship between laryngeal carcinoma and LNM and to improve the model’s performance and generalizability. This study presents a novel ITH-aware DL framework that integrates adaptive ITH mapping with deep learning to preoperatively predict LNM in LSCC. By combining radiomics, 3D DL, and ITH-aware DL model within a stacking ensemble architecture, we achieved improved predictive performance for LNM. Both the ITH-aware DL and stacking models stratified patients into high-risk and low-risk groups with different OS outcomes, reflecting the intrinsic biological link between LNM propensity and tumor aggressiveness. Our findings demonstrate that non-invasive, imaging-based prediction of LNM probability can enable preoperative risk stratification and potentially guide clinical decision making. This approach provides a robust tool for personalized surgical planning and prognostic evaluation in LSCC, supporting the optimization of treatment strategies. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1