Multimodal deep learning integration of clinical, dosiomics, and radiomics features to predict toxicity in breast cancer patients undergoing hypofractionated radiotherapy.
Authors
Affiliations (5)
Affiliations (5)
- Department of Biomedical Engineering, National Taiwan University, Taipei, 106, Taiwan; Division of Radiation Oncology, Department of Oncology, National Taiwan University Hospital Yunlin Branch, Yunlin County, 632, Taiwan; Division of Radiation Oncology, Department of Oncology, National Taiwan University Hospital, College of Medicine, National Taiwan University, Taipei, 100, Taiwan.
- Department of Electrical Engineering, National Yunlin University of Science and Technology, Yunlin, 640, Taiwan.
- Division of Radiation Oncology, Department of Oncology, National Taiwan University Hospital Yunlin Branch, Yunlin County, 632, Taiwan.
- Department of Biomedical Engineering, National Taiwan University, Taipei, 106, Taiwan; Division of Radiation Oncology, Department of Oncology, National Taiwan University Hospital, College of Medicine, National Taiwan University, Taipei, 100, Taiwan; Department of Radiation Oncology, National Taiwan University Cancer Center, National Taiwan University Hospital and National Taiwan University College of Medicine, Taipei, 106, Taiwan.
- Department of Electrical Engineering, National Yunlin University of Science and Technology, Yunlin, 640, Taiwan. Electronic address: [email protected].
Abstract
Radiation dermatitis (RD) and superficial soft tissue fibrosis are common toxicities among the patients with breast cancer receiving hypofractionated radiotherapy (HFRT). Early identification of high-risk patients could enable proactive supportive care and optimized treatment planning. We developed and evaluated multimodal machine learning integrating clinical, radiomic, dosiomic, and deep learning (DS) features and models to predict acute dermatitis after HFRT delivered with intensity-modulated radiation therapy (IMRT). Fibrosis analysis was exploratory because of the limited number of events. This pre-register single-center cohort with retrospective analysis examined early-stage breast cancer patients treated with IMRT-HFRT between 2017 and 2022. Radiomic features were extracted from planning CT images using the planning target volume as the region of interest. Dosiomic features were derived from 3D dose distributions using 5 Gy dose-binning for the whole breast and tumor bed. DS features were obtained from DenseNet-121 applied to resampled CT and dose images. Two analytical frameworks were compared: (1) conventional models using clinical, radiomic, and dosiomic features; and (2) an expanded framework adding DS features, CLAHE-based image enhancement, and SMOTE class rebalancing. Models (Random Forest, XGBoost, LightGBM, and DS-LightGBM) and SMOTE were trained using 5-fold cross-validation within the training set and tested on independent data. SHAP-based feature selection was used to improve performance and reduce dimensionality. There were 160 patients analyzed in the end. RD Grades 0-3 and binary dermatitis (Grades 0-1 vs 2-3) occurred in 15/81/48/16 cases and 40% with Grade ≥2, respectively. Grade 1 fibrosis was observed in 5 patients. The DS-LightGBM model achieved the best performance across endpoints: multiclass prediction (test accuracy 0.87, ROC-AUC 0.95), binary classification (accuracy 0.89, ROC-AUC 0.88-0.90), and fibrosis (accuracy 0.97, ROC-AUC 0.85-0.89). SHAP analysis showed that high-dose dosiomic subvolumes and DS features were predominant predictors. Fibrosis modeling was considered exploratory because of the rare-event setting. Multimodal integration of clinical, radiomic, dosiomic, and DS features in DS-LightGBM showed promising internal performance for RD prediction after HFRT in early-stage breast cancer. These models may support personalized planning and early toxicity mitigation and improve the patients' quality of life. Further external validation is warranted before clinical use. Fibrosis modeling should be interpreted as exploratory because only 5 events were observed.