Dose-Guided Hybrid AI Model with Deep and Handcrafted Radiomics for Explainable Radiation Dermatitis Prediction in Breast Cancer VMAT

Abstract

Purpose: To improve the prediction accuracy of radiation dermatitis (RD) in breast cancer patients undergoing volumetric modulated arc therapy (VMAT), we developed a hybrid artificial intelligence (AI) model that integrates deep learning radiomics (DLR), handcrafted radiomics (HCR), clinical features, and dose-volume histogram (DVH) parameters, aiming to enhance the early identification of high-risk individuals and support personalized prevention strategies.

Methods: A retrospective cohort of 156 breast cancer patients treated with VMAT at Kaohsiung Veterans General Hospital (2018-2023) was analyzed; 148 patients were eligible after exclusions, with RD graded according to the RTOG criteria. Clinical variables and 12 DVH indices were collected, while HCR features were extracted via PyRadiomics. DLR features were derived from a pretrained VGG16 network across four input designs: original CT images (DLR_Original), a 5 mm subcutaneous region (DLR_Skin5mm), a planning target volume with a 100% prescription dose (DLR_PTV100%), and a subcutaneous region receiving ≥ 5 Gy (DLR_V5Gy). The features were preselected via ANOVA (p < 0.05), followed by Boruta-SHAP refinement across 11 feature sets. Predictive models were built via logistic regression, random forest, gradient boosting decision tree, and stacking ensemble (SE) methods. Explainability was assessed via SHapley Additive exPlanations (SHAPs) and gradient-weighted class activation mapping (Grad-CAM).

Results: Among the 148 patients, 49 (33%) developed Grade ≥ 2 RD. The DLR models outperformed the HCR models (AUC = 0.72 vs. 0.66). The best performance was achieved with DLR_V5Gy + clinical + DVH features, yielding an AUC = 0.76, recall = 0.68, and F1 score = 0.60. SE consistently surpassed single classifiers. SHAP identified convolutional DLR features as the strongest predictors, whereas Grad-CAM focused attention on subcutaneous high-dose regions, which was consistent with the clinical RD distribution.

Conclusions: The proposed hybrid AI framework, which integrates DLR, clinical, and DVH features, provides accurate and explainable predictions of Grade ≥ 2 RD after VMAT in breast cancer patients. By combining ensemble learning with XAI methods, the model offers reliable high-risk stratification and potential clinical utility for personalized treatment planning.

Keywords: breast cancer; deep learning radiomics; ensemble learning; explainable artificial intelligence; radiation dermatitis; volumetric modulated arc therapy.

Figure 1

Workflow of the proposed hybrid ensemble artificial intelligence (AI) framework for predicting radiation dermatitis in breast cancer patients. Abbreviation: DVH, Dose-Volume Histogram; CT, Computed Tomography; HCR, Handcrafted Radiomics; DLR, Deep Learning Radiomics; ANOVA, Analysis of Variance; LASSO, Least Absolute Shrinkage and Selection Operator; AI, Artificial Intelligence; ML, Machine Learning; LR, Logistic Regression; RF, Random Forest; GBDT, Gradient Boosting Decision Tree; AUC, Area Under the ROC Curve; ROC, Receiver Operating Characteristic; ACC, Accuracy; NPV, Negative Predictive Value; PPV, Positive Predictive Value; Grad-CAM, Gradient-weighted Class Activation Mapping; Valid, Validation set; SHAP, SHapley Additive exPlanations; SMOTE–ENN, Synthetic Minority Over-sampling Technique combined with Edited Nearest Neighbors.

Figure 2

Illustration of DLR input image sources: (a) DLR_Original, (b) DLR_Skin5mm, (c) DLR_PTV100%, (d) DLR_V5Gy. Abbreviation: DLR, Deep Learning Radiomics; ROI, Region of Interest. Colored blocks below each panel represent visual labels for the four DLR input categories and do not convey dose or intensity information.

Figure 3

Grad-CAM visualization of DLR features across VGG16 convolutional layers. (a) DLR_Original: unmasked thoracic CT image. (b) DLR_Skin5mm: contoured 5-mm subcutaneous skin layer. (c) DLR_PTV100%: overlap between the 100% prescription-dose region and the 5-mm skin layer. (d) DLR_V5Gy: intersection of the 5-mm skin layer with the region receiving ≥ 5 Gy. Warm colors indicate higher model attention, whereas cool colors indicate lower attention. Block and font colors are used for visual grouping only and do not convey quantitative information. Abbreviations: DLR, Deep Learning Radiomics; VGG16, Visual Geometry Group 16; Grad-CAM, Gradient-weighted Class Activation Mapping.

Figure 4

(a) Effect of ROI input design on the spatial attention patterns of DLR models across different deep-learning feature inputs. (b) Spatial correspondence between Grad-CAM hot-spots and high-risk dose regions associated with radiation dermatitis (RD). The Grad-CAM heatmaps are color-coded from blue to red, where blue indicates low model attention, red indicates high attention, and intermediate green–yellow colors represent moderate attention. Block colors beneath the panels are used only for visual grouping and do not represent dose or image intensity. Abbreviations: DLR, Deep Learning Radiomics; ROI, Region of Interest; PTV, Planning Target Volume; Grad-CAM, Gradient-weighted Class Activation Mapping.

Figure 5

SHAP summary plot for Feature Combination 11 (Clinical + DVH + DLR_V5Gy). Each point represents a SHAP value for an individual patient, illustrating both the magnitude and direction of the feature’s influence on the model output. Features are ranked by overall importance from top to bottom. The color scale represents normalized feature values, where red indicates high values and blue indicates low values. Positive SHAP values indicate increased predicted risk of radiation dermatitis; negative values indicate reduced risk. Abbreviations: DLR, Deep Learning Radiomics; DVH, Dose–Volume Histogram; SHAP, SHapley Additive exPlanations.