Background: Response to neoadjuvant therapy in locally advanced rectal cancer (LARC) is heterogeneous, and early identification of non-responders may help optimize treatment strategies and reduce unnecessary toxicity. This study aimed to develop and internally validate a machine learning model based on radiomic features extracted from baseline magnetic resonance imaging (MRI) to predict treatment response defined according to MRI tumor regression grade (mrTRG) at restaging MRI. Methods: In this retrospective single-center study, 86 patients with histologically confirmed LARC who underwent baseline and restaging MRI, neoadjuvant therapy, and surgery were included. Primary tumors were manually segmented on oblique axial T2-weighted images. A total of 107 radiomic features were extracted using PyRadiomics (vrs 3.0.1), with and without N4 bias field correction. Feature selection was performed using LASSO, followed by elastic net–regularized logistic regression. Model performance was evaluated using repeated stratified 5-fold cross-validation (20 repetitions). Treatment response was defined according to MRI tumor regression grade (mrTRG) at restaging, dichotomized into responders (mrTRG ≤ 2) and non-responders (mrTRG ≥ 3). Results: The model achieved a mean area under the receiver operating characteristic curve (AUC-ROC) of 0.73, with an accuracy of 72.5%, sensitivity of 79.2%, and specificity of 50%. Conclusions: Baseline MRI-based radiomics shows potential for identifying patients at higher risk of non-response to neoadjuvant therapy in LARC. However, limited specificity and the absence of external validation restrict immediate clinical applicability. Further validation in larger multicenter cohorts and integration with clinical variables are warranted to improve model robustness and generalizability.
Ambrosini et al. (2026) studied this question.