A spatio-temporal dual-view deep learning model for early prediction of grade ≥ 2 radiation dermatitis

To develop and internally evaluate a deep learning model for early prediction of grade ≥ 2 radiation dermatitis in breast cancer patients undergoing postoperative radiotherapy. This prospective single-center cohort study included 626 predominantly Chinese patients receiving adjuvant radiotherapy. Standardized skin photographs of the breast/chest wall and axillary regions were acquired at multiple time points during treatment. A dual-view deep learning framework was constructed, combining spatial feature extraction (EfficientNetV2-L) with temporal modeling using a gated recurrent unit (GRU). The primary endpoint was ≥ grade 2 radiation dermatitis according to RTOG criteria. Model performance was internally evaluated using patient-level five-fold cross-validation. In addition, dynamic prediction analyses were performed using image sequences from different early treatment windows. The model achieved an AUC of 0.9129 (95% CI: 0.8893–0.9345), with a sensitivity of 0.8556 (95% CI: 0.8105–0.8946) and specificity of 0.8386 (95% CI: 0.7994–0.8758). Ablation analyses confirmed the contribution of both spatial and temporal components, with a marked performance drop when temporal modeling was removed (AUC 0.6997). In dynamic analyses, the model reached an AUC of approximately 0.80 using only the first three weeks of data, allowing early risk stratification before the period when clinically significant dermatitis typically becomes more apparent. Performance peaked at week 5 (AUC 0.8931). Decision curve analysis demonstrated consistent net benefit across a range of clinically relevant thresholds. A spatio-temporal deep learning approach enables early prediction of ≥ grade 2 radiation dermatitis as early as week 3 during radiotherapy. This framework may support proactive risk assessment and facilitate individualized management of radiation-induced skin reactions. Further multicenter external validation is warranted before broader clinical application.