An Interpretable Multimodality Deep Learning Model for Prognostic Prediction and Risk Stratification In Stage III-IVA Squamous Cervical Cancer Using 18f-FDG PET/CT
Abstract
Purpose
High-risk locally advanced cervical cancer (HR-LACC) remains challenging due to its heterogeneous prognosis and high recurrence rates, underscoring the need for precise and personalized treatment strategies. This study aims to develop and validate an interpretable multimodal prognostic model by integrating deep learning-based radiomic features (DLR) from 18F-FDG PET/CT, traditional handcrafted radiomic features (HCR), and clinical data to predict progression-free survival (PFS) in HR-LACC patients undergoing concurrent chemoradiotherapy (CCRT).
Methods
This retrospective study involved 195 HR-LACC patients from two institutions. Clinical, HCR, and DLR features were extracted from pre-treatment 18F-FDG PET/CT scans. A UNet3D model was used for tumor segmentation, and LASSO regression was applied for feature selection. Multimodal fusion of clinical, HCR, and DLR features was employed to construct a Random Forest (RF) model for 3-year PFS prediction. Model performance was evaluated using AUC, C-index, and Kaplan-Meier survival analysis, with external validation using an independent cohort. Finally, the SHAP method was employed to quantify the contributions of the Clinical, HCR, and DLR features in the model's predictions, providing explanations for both global and local conditions.
Results
The final multimodality model, COMB-Rad, integrating clinical, HCR, and DLR features, demonstrated excellent performance with an AUC of 0.844 ± 0.028 and a C-index of 0.759 in the internal validation cohort. The model effectively stratified patients into high- and low-risk groups, with significant differences in 3-year PFS (9.0% vs. 81.3%, p < 0.0001). The external validation cohort showed consistent results, highlighting the model's robust generalizability.
Conclusion
The multimodality prognostic model COMB-Rad provides an effective tool for risk stratification in HR-LACC patients. It demonstrates stable risk-stratification capabilities in multi-center validation, providing high-precision support for personalized treatment strategies and improved clinical decision-making.