A Machine Learning Model Based on MRI Radiomics to Predict Response to Chemoradiation among Patients with Rectal Cancer

Background. In the context of increasing use of rectum sparing protocols for rectal cancer treatment, identifying patients with a pathological complete response (pCR) to preoperative chemoradiotherapy (pCRT) remains challenging. Objective. This study utilizes advanced MRI and a robust radiomic-based machine-learning approach to predict pCRT response in rectal cancer patients extracting features from pre-treatment staging MRI. We investigate the potential of radiomic features in staging MRI in identifying pCR in locally advanced rectal cancers. Materials and Methods. Pre-treatment staging MRI studies from 102 subjects were collected, divided in a training (n=72) and validation (n=30) cohort. In the training cohort, 52 patients were classified as not responding to pCRT and 20 patients as pCR, based on the histological diagnosis from total mesorectal excision (reference standard). This dataset was employed for training and cross-validation of different machine-learning models. A comprehensive radiomic approach was applied, hypothesizing that radiomic features could effectively capture disease heterogeneity explaining the different treatment response between the two groups. The model was subsequently tested in the external validation cohort where 26 patients were classified as not responding and 4 as pCR by the reference standard. Results. Three machine-learning models were developed, achieving a best-performing model with a Receiver Operating Characteristic Area Under the Curve (ROC-AUC) of 73% and an accuracy of 70%. Sensitivity and Positive Predictive Value (PPV) reached 78% and 80%, respectively. Testing on the validation cohort showed sensitivity of 81%, specificity of 75%, and accuracy of 80%. Conclusion. This study demonstrates the potential of a radiomic-based machine-learning approach to predict treatment response in rectal cancer patients using staging MRI. The top-performing model shows promising discriminative capabilities, highlighting the importance of integrating advanced imaging and computational methodologies in personalized rectal cancer management.