Enhancing the prediction accuracy of pathological downstaging in locally advanced rectal cancer using deep learning models with preoperative MRI and clinicopathological data.

PURPOSE: Conventional magnetic resonance imaging (MRI) for locally advanced rectal cancer (LARC) involves challenges in evaluating and predicting the preoperative response to neoadjuvant chemoradiotherapy (CRT). Deep learning (DL) methods incorporating MRI are widely utilized for cancer diagnosis and outcome prediction. This study aims to develop and validate DL models based on preoperative T2-weighted MR images combined with radiological and clinicopathological data, and to assess their performance in predicting pathological T (pT)-downstaging after CRT. METHODS: This retrospective study included 318 patients with histopathologically confirmed LARC, randomly assigned to a training set (n = 223) or internal test set (n = 95). An additional 88 patients comprised the external test set. DL models integrating T2-weighted MR images with radiological and clinicopathological characteristics were constructed to predict pT-downstaging. Model performance was evaluated via receiver operating characteristic (ROC) analysis; DeLong's test was used to compare ROC curves. RESULTS: The combined models (clinicopathological + T2-weighted MR images, MRI characteristics + T2-weighted MR images, and the all-combined model) demonstrated superior diagnostic performance in both internal and external test sets, with areas under the curve (AUCs) ranging from 0.800 to 0.817 and 0.802 to 0.810, respectively. AUCs of the T2-weighted MRI-only model and the combined models significantly differed (P < 0.05); no significant differences were observed among the three combined models. CONCLUSION: The proposed combined DL models demonstrated strong predictive performance for pT-downstaging after neoadjuvant CRT in LARC, with the MRI characteristics + T2-weighted MR images model also performing well without clinicopathological features.