Multi-modal predictive modeling in prostate cancer radiotherapy: A radiomics and dosiomics approach

This study aims to create a predictive model for evaluating how prostate cancer patients respond to radiotherapy. The model combines radiomic, dosiomic, and clinical data to improve treatment planning and distinguish between responders and non-responders. The study included 387 prostate cancer patients who underwent radiotherapy. Pre-treatment imaging consisted of CT scans and MRI (T2-weighted and diffusion-weighted imaging), from which 215 radiomic and dosiomic features were analyzed. Predictive metrics like Apparent Diffusion Coefficient (ADC) values and changes in signal intensity before and after treatment were evaluated. Feature selection was performed using both Recursive Feature Elimination (RFE) and Principal Component Analysis (PCA). RFE, a supervised method, was used to identify the most predictive features by recursively removing the least important ones based on model performance. PCA, an unsupervised dimensionality reduction technique, was applied to capture variance and address multicollinearity in high-dimensional feature sets. Machine learning models, including XGBoost, CatBoost, and Stacking, were trained and validated through stratified k-fold cross-validation. Performance metrics like Area Under the Curve (AUC), accuracy, sensitivity, and specificity were used to assess model effectiveness. Integrating all features yielded the highest predictive performance. The Stacking model (using RFE-selected features) achieved an AUC of 88.14 %, an accuracy of 87.17 %, and a sensitivity of 87.18 % in testing. Dosiomic features alone were highly predictive, achieving an AUC of 78.33 %. Radiomic features also played a crucial role, particularly texture and morphology metrics. Responders demonstrated significant increases in ADC values (from 586 mm 2 /s to 1050 mm 2 /s) and notable decreases in signal intensity (from 682 to 100). In contrast, non-responders showed minimal changes in these metrics. These differences proved effective in differentiating patient outcomes. This study highlights the importance of integrating radiomic, dosiomic, and clinical data to predict treatment responses in prostate cancer radiotherapy accurately.