Leveraging Inter-Radiomic Feature Relationships for Enhanced Prediction of Distant Metastasis and Characterization of Heterogeneity in Head and Neck Cancer.
Aug 2026· International Journal of Radiation Oncology, Biology, Physics· 0 citations· 39 references
Medicine
TL;DR
Modeling inter-radiomic feature relationships using the OmicsMap representation substantially improves CT-based prediction of DMFS and characterization of tumor heterogeneity in HN cancer, supporting precision risk stratification in clinical oncology.
Abstract
Purpose
Distant metastasis remains a major cause of treatment failure in head and neck (HN) cancer, highlighting the need for more accurate early risk stratification. This study developed and validated a deep radiomics framework to characterize tumor heterogeneity from pre-treatment CT images and improve prediction of distant metastasis-free survival (DMFS).
Materials And Methods
This multicenter study included 3,421 HN cancer patients from four cohorts across twelve institutions (RADCURE, HN1, HN-PET-CT and TCGA-HNSC). Radiomics features were extracted from primary tumors and transformed into OmicsMaps, a structured representation that spatially organizes inter-feature relationships to facilitate learning of complex prognostic patterns. A convolutional neural network was trained to derive prognostic signatures, which were integrated with key clinical variables to construct an OmicsMap-clinical fusion model for patient risk stratification. Model performance was assessed using the concordance index (C-index) and time-dependent area under the receiver operating characteristic curve (AUC) in the RADCURE, HN1, and HN-PET-CT cohorts. Radiogenomic analyses using RNA-seq data were conducted in the TCGA-HNSC cohort to investigate biological characteristics associated with the imaging-defined risk groups.
Results
The OmicsMap achieved C-index values of 0.742, 0.768, and 0.671 in the RADCURE, HN1, and HN-PET-CT cohorts, outperforming conventional radiomics approach by 5.84-6.37%. Incorporating clinical variables further improved generalizability, yielding C-index of 0.864 (HN1) and 0.730 (HN-PET-CT), with time-dependent AUCs of 0.727-0.895. The fusion model consistently stratified patients into distinct high- and low-risk groups for both DMFS and overall survival across cohorts (P < 0.01). Radiogenomic analyses revealed enrichment of immune-related pathways in the low-risk group, whereas the high-risk group exhibited a more aggressive phenotype enriched for proliferation, hypoxia, and epithelial-mesenchymal transition pathways, along with a fibrosis-prone tumor microenvironment characterized by extracellular matrix remodeling.
Conclusion
Modeling inter-radiomic feature relationships using the OmicsMap representation substantially improves CT-based prediction of DMFS and characterization of tumor heterogeneity in HN cancer, supporting precision risk stratification in clinical oncology.
RATIONALE AND OBJECTIVES
Breast cancer carries a substantial long-term risk of bone metastasis, which marks progression to incurable disease and severely compromises patient quality of life. Conventional clinicopathological factors provide limited individualized prognostic resolution, highlighting the need for noninvasive biomarkers capable of capturing tumor biology relevant to bone-specific metastatic progression.
MATERIALS AND METHODS
We systematically compared two commonly used intratumoral heterogeneity (ITH) modeling strategies: feature-level heterogeneity encoding based on tumor radiomics features complexity (ITH1) and habitat-based regional aggregation derived from supervoxel partitioning (ITH2). Prognostic performance for bone metastasis-free survival (BMFS) was also evaluated by global tumor region (GTR) radiomics and vision transformer (ViT)-derived deep learning features. Multidimensional models integrating imaging-derived scores and clinicoradiological variables were constructed and interpreted using SHapley Additive exPlanations (SHAP).
RESULTS
ITH1 consistently outperformed ITH2 in prognostic performance and generalizability, indicating greater robustness in capturing tumor heterogeneity relevant to bone metastasis. Integration of the ITH1 score with GTR radiomics and ViT scores significantly improved BMFS prediction compared with any single modality. The final integrated model incorporating clinicoradiological variables achieved high and stable discrimination across cohorts, with concordance index values ranging from 0.878 to 0.908. SHAP analysis revealed that the ITH1 score contributed most prominently to bone metastasis risk prediction, underscoring the critical role of tumor heterogeneity in bone-specific metastatic risk.
CONCLUSION
Feature-level ITH encoding provides a more informative and robust representation of ITH than habitat-based modeling for predicting breast cancer bone metastasis. Multidimensional integration of ITH, GTR radiomics, and ViT-derived features enables complementary tumor characterization and substantially enhances individualized risk stratification, supporting its potential utility in precision oncology.
Liwei Sun, Weizhi Zhang, Zhengtong Wang et al.· Academic Radiology· 0 citations
BACKGROUND
Accurate prognostic stratification following concurrent chemoradiotherapy (CCRT) for cervical cancer remains challenging. Existing approaches often rely on contrast-enhanced magnetic resonance imaging (MRI), which limits applicability in patients with renal dysfunction and in resourcelimited settings. We aimed to develop and validate a non-contrast MRI-based prognosis prediction model combined with clinical predictors to estimate recurrence risk and progression-free survival (PFS).
MATERIALS AND METHODS
A total of 145 cervical cancer patients from two centers were retrospectively analyzed. The model-building cohort (n=114) was used to extract radiomics features from pre-treatment MRI, and an extreme gradient boosting classifier was developed. Clinical predictors were identified through univariate Cox and multivariate logistic regression analyses. A nomogram integrating radiomics and clinical variables was then constructed. External validation was performed in an independent cohort (n=31). Model performance was assessed by discrimination, calibration, and its ability to stratify patients into low- and high-risk groups for predicting 5-year PFS.
RESULTS
The proposed method categorized patients into well-defined risk groups. Survival results varied greatly among groups. Kaplan-Meier analysis indicated that patients with high risk had a worse 5-year PFS compared to those with low risk. This result was observed in both the construction cohort (P < 0.001) and the external validation cohort (P=0.012). The prognostic performance was comparable to that of contrast-enhanced MRI-based models. A significant point to note is that no gadolinium administration was necessary.
DISCUSSION
This study develops a radiomics nomogram that relies on non-contrast MRI. The model offers strong and tailored prognostic information. Imaging features and clinical predictors combined, which improved predictive accuracy. The approach increases accessibility, particularly for patients contraindicated for contrast or in low-resource settings.
CONCLUSION
We developed and validated a non-contrast MRI radiomics nomogram that accurately predicts recurrence and 5-year PFS following CCRT in cervical cancer. The model offers a clinically applicable, cost-effective tool for personalized risk stratification.
Wei Wang, Xiao-ting Li, Yue-Qi Jiang et al.· Current medical imaging· 0 citations
Lymphovascular space invasion (LVSI) is an important pathological feature associated with tumor aggressiveness and adverse prognosis in cervical cancer. However, reliable preoperative prediction of LVSI remains a challenge. This study aimed to develop and validate an MRI-based radiomics model incorporating intratumoral and peritumoral features for LVSI prediction and systematically evaluate the optimal peritumoral extent.
In this single-center study comprising both retrospective and prospective cohorts, 204 patients with pathologically confirmed cervical cancer were randomly divided into a training cohort (n = 142) and a test cohort (n = 62). Radiomics features were extracted from the intratumoral and peritumoral regions with expansion distances of 1, 3, and 5 mm. Feature selection was performed using intraclass correlation coefficient (ICC) analysis, minimum redundancy maximum relevance (mRMR), and least absolute shrinkage and selection operator (LASSO) analysis. Six machine learning algorithms, including logistic regression [LR], support vector machine, random forest, ExtraTrees, LightGBM, and multilayer perceptron, were used to construct the predictive models. The model performance was evaluated using receiver operating characteristic analysis, calibration curves, and decision curve analysis. SHapley Additive Explanations (SHAP) were applied to interpret the optimal models.
Among all ROI configurations, the Intra+P1 model demonstrated the best overall performance, particularly when it was combined with LR. The LR-based Intra+P1 model achieved an AUC of 0.866 (95% CI: 0.807–0.926) in the training cohort and 0.843 (95% CI: 0.731–0.955) in the test cohort respectively. Increasing the peritumoral expansion from 1 mm to 3 mm or 5 mm did not further improve predictive performance. The addition of clinical variables (tumor diameter and SCC level) did not provide a significant incremental predictive value. Calibration and decision curve analyses demonstrated good agreement and favorable clinical utility of the model. SHAP analysis showed that both intratumoral and peritumoral features contributed substantially to the model prediction, with texture features playing a dominant role.
This interpretable MRI-based radiomics model facilitates accurate preoperative prediction of LVSI in cervical cancer. A narrowly defined 1-mm peritumoral region provides the most informative complementary information, underscoring the importance of the tumor-invading front. This approach offers a noninvasive tool for risk stratification and may support individualized treatment decision-making.
Xian-Yan Wu, Chuanfang Xu, Sha Shi et al.· Frontiers in Oncology· 0 citations
PURPOSE
Accurate prediction of distant metastasis (DM) and locoregional recurrence (LR) in head and neck cancer (HNC) patients following radiotherapy is critical for individualized treatment planning. This study aimed to develop and validate an explainable deep learning model integrating anatomical and dosimetric information for HNC DM and LR risk stratification.
METHODS AND MATERIALS
237 HNC patients treated with definitive radiotherapy were collected and divided into training, internal, and external validation cohorts. A dual-task deep learning model based on 3D squeeze-and-excitation residual networks was constructed to predict DM and LR risk using CT images, 3D dose distributions, and gross tumor volume (GTV) masks. Model performance was evaluated by concordance index (C-index), time-dependent ROC, and decision curve analysis. Interpretability was enhanced using Grad-CAM and a novel Activation Volume Histogram (AVH) method to quantify attention patterns across spatial-dosimetric subregions.
RESULTS
The model achieved C-indices of 0.92 and 0.74 for DM and LR, respectively, outperforming radiomic models. Grad-CAM visualizations revealed distinct activation patterns aligned with recurrence sites- lymphatic regions for DM and tumor zones for LR. AVH analysis identified GTV extended by 3mm receiving dose ≥ 65 Gy and GTV receiving dose ≥ 65 Gy as the most discriminative subregions, showing significant differences in activation distributions between high- and low-risk groups, particularly within the 0.4-0.5 intensity range.
CONCLUSION
This dual-task model enables accurate DM risk prediction and showed potential for LR risk prediction in HNC, providing spatial-dosimetric information to support risk-adaptive radiotherapy. Further extensive multicenter validation is warranted to confirm its clinical applicability.
Huali Li, Jiajun Cai, Xuan-Ru Zhou et al.· International Journal of Rad...· 0 citations
Objectives To develop and validate a preoperative multimodal model that predicts a high-risk recurrence phenotype indicative of poor disease-free survival (DFS), in order to stratify patients with stage IA non-small cell lung cancer (NSCLC). Materials and methods This retrospective multicenter study enrolled 342 stage IA NSCLC patients from three independent centers. The high-risk recurrence phenotype (indicative of poor DFS) was defined by postoperative pathology as the presence of spread through air spaces (STAS), lymphovascular invasion (LVI), a predominant solid/micropapillary/complex glandular pattern, or a > 5% solid/micropapillary component. Deep learning features were extracted from preoperative biopsy whole-slide images (WSI) using the UNI foundation model, and CT morphological and textural features were extracted from preoperative chest CT. An early-fusion multimodal model integrating clinical, radiomics, and pathomics features was developed and evaluated with five-fold cross-validation. The Kaplan–Meier method with log-rank tests was used to assess associations between the model-predicted risk and DFS. Logistic regression identified clinical predictors of high-risk pathology. Model interpretability and clinical utility were examined with SHapley Additive exPlanations (SHAP) and calibration analysis, respectively. Results The multimodal model achieved higher discriminative performance than each single-modality model in both the internal and external test sets. In the internal test set, it yielded an AUC of 0.76 (95% CI 0.58-0.90); in external validation, AUCs were 0.69 (95% CI 0.55-0.82) in Center 2 and 0.86 (95% CI 0.74-0.96) in Center 3. Multivariable analysis identified solid tumor density as the only independent predictor (OR = 5.13, 95% CI 1.53-17.24; P = 0.008). Model-stratified high-risk patients showed significantly inferior DFS in both the training (2-year DFS 78% vs. 99%; log-rank P < 0.0001) and external (3-year DFS 86% vs. 100%; log-rank P = 0.003) cohorts. Conclusion The multimodal model consistently predicted the high-risk recurrence phenotype across multiple centers. This phenotype may serve as a pragmatic indicator of poor DFS to guide earlier individualized treatment decisions, including adjuvant therapy and intensified surveillance, in stage IA NSCLC.
L. Lai, Yanji Wang, Jiangle Jiang et al.· Frontiers in Immunology· 0 citations
Background Despite substantial therapeutic progress, distant metastasis (DM) persists as the leading cause of treatment failure and a critical driver of poor survival in nasopharyngeal carcinoma (NPC). Existing prognostic tools remain limited in accuracy and clinical utility, underscoring the need for more robust and actionable biomarkers. The convergence of artificial intelligence (AI) and digital pathology offers a promising avenue by enabling the extraction of prognostic signatures directly from routinely available histopathology. In this study, we developed and validated a deep learning (DL) model based on whole slide images (WSIs) to predict post-treatment DM risk in patients with locally advanced NPC. Methods We developed the Pathology-based distant metastasis risk stratification model (P-MES) using WSIs from 147 NPC patients. WSIs were encoded into patch-level embeddings using the pretrained UNI feature extractor, followed by multiple instance learning (MIL) aggregation to generate slide-level representations for predicting DM. To enhance interpretability, attention-based heatmaps were generated, and complementary morphologic features extracted by CellProfiler were used to elucidate histopathologic patterns associated with metastatic risk. Results P-MES achieved excellent predictive performance in the training, validation, and internal test sets (AUCs: 0.990, 0.887, and 0.949), demonstrating substantial discrimination, calibration, and clinical net benefit. CellProfiler analysis further identified ten morphology-derived features associated with DM. Conclusion P-MES is an automated, pathology-based framework for predicting DM in NPC, providing pathological insights and supporting personalized clinical management.
Liuling Wang, Jiaxin Lin, Mengting Xu et al.· Digital Health· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
