Multiparametric MRI-Based Deep Learning Models for Preoperative Prediction of Tumor Deposits in Rectal Cancer and Prognostic Outcome.

Rationale And Objectives: To investigate the predictive value of a deep learning model based on multiparametric MRI (mpMRI) for tumor deposit (TD) in rectal cancer (RC) patients and to analyze their prognosis.

Materials And Methods: Data from 529 RC patients who underwent radical surgery at two centers were retrospectively collected. 379 patients from center one were randomly divided into a training set (n = 265) and an internal validation (invad) set (n = 114) in a 7:3 ratio. 150 patients from center two were included in the external validation (exvad) set. Univariate and multivariate analyses were performed to identify independent clinical predictors and to construct a clinical model. Preoperative mpMRI images were utilized to extract deep features through the ResNet-101 model. Following feature selection, a deep learning model was developed. A nomogram was created by combining the clinical model with the deep learning model. The clinical applicability of each model was assessed using ROC curves, decision curve analysis (DCA), clinical impact curves (CIC), and deLong test. Kaplan-Meier survival analysis was conducted to evaluate prognostic outcome among patients.

Results: Among the 529 patients, 142 (26.8%) were TD positive. In the training set, clinical model was constructed based on clinical independent predictors (cT and cN). 30 deep features were selected to calculate the deep learning radscore (DLRS) and develop the deep learning (DL) model. The AUC values for the clinical model were 0.724, 0.836, and 0.763 in the training set, invad set, and exvad set, respectively. The AUC values for the DL model were 0.903, 0.853, and 0.874, respectively. The nomogram achieved higher AUC values of 0.925, 0.919, and 0.9, respectively. The DeLong test indicated that the predictive performance of the nomogram was superior to both the DL model and the clinical model in training and invad sets. Kaplan-Meier survival analysis showed that both the DL model and the nomogram effectively stratified patients into high-risk and low-risk groups for 3-year DFS (p < 0.05).

Conclusion: The nomogram, which integrates mpMRI-based deep radiomic features and clinical characteristics, effectively predicts preoperative TD status in RC. Both the DL model and the nomogram can effectively stratify patients' 3-year DFS risk.