A novel deep learning segmentation and multimodal radiomics approach for radiation proctitis prediction in cervical cancer.
Authors
Affiliations (9)
Affiliations (9)
- Department of Oncology, The First Affiliated Hospital of Yangtze University, Jingzhou, Hubei, China.
- Hubei Provincial Clinical Research Center for Personalized Cancer Diagnosis and Therapy, Jingzhou, China.
- Department of Oncology, The First Affiliated Hospital of Yangtze University, Jingzhou, Hubei, China. [email protected].
- Hubei Provincial Clinical Research Center for Personalized Cancer Diagnosis and Therapy, Jingzhou, China. [email protected].
- Hubei Province Key Laboratory of Precision Radiation Oncology, Wuhan, 430022, China. [email protected].
- Department of Oncology, The First Affiliated Hospital of Yangtze University, Jingzhou, Hubei, China. [email protected].
- Hubei Provincial Clinical Research Center for Personalized Cancer Diagnosis and Therapy, Jingzhou, China. [email protected].
- Hubei Province Key Laboratory of Precision Radiation Oncology, Wuhan, 430022, China. [email protected].
- Department of Oncology, The First Affiliated Hospital of Yangtze University, Jingzhou, Hubei, China. [email protected].
Abstract
This study aimed to develop and validate a novel interpretable framework integrating a deep segmentation network with multimodal radiomics to improve the prediction performance of radiation proctitis (RP) in cervical cancer patients undergoing radiotherapy. This retrospective study included 650 cervical cancer patients, who were divided into training and internal device test sets, and independent device test set, based on different computed tomography (CT) scanners. A Mamba-enhanced deep learning network was constructed to enhance global modeling capability in pelvic anatomical structures, providing reliable regions of interest segmentation for radiomics and adaptive radiotherapy. In multimodal radiomics approach, foundation model features were integrated with radiomics and clinical data to mitigate inherent bias in training data, thereby enhancing the prediction of RP. Shapley Additive exPlanations (SHAP) technology was employed to explain the contributions of multimodal features to the prediction performance. The segmentation model achieved dice similarity coefficients (DSC) of 90.76% and 85.33% on the internal and independent device test sets, outperforming existing models. Ablation study confirmed the contribution of each module to segmentation performance. The multimodal radiomics model demonstrated robust performance in predicting RP, with area under the curve (AUC) values of 0.882 and 0.865 on the internal and independent device test sets, surpassing single-feature models. Additionally, SHAP analysis revealed the interactions between key features, including dosimetric and foundation model features, which contributed to improve decision interpretability. This study significantly improved the robustness and accuracy of RP prediction by integrating the deep segmentation model and multimodal radiomics, providing a promising clinical applications framework for individualized radiotherapy planning. Not applicable.