To explore the application value of deep learning (DL)-based segmentation and classification models using MRI for preoperative TN staging of rectal cancer.

A retrospective analysis was conducted on clinical, pathological, and pre-treatment MRI data from patients with postoperatively pathologically confirmed rectal cancer at the First Affiliated Hospital of Naval Medical University between January 2019 and December 2023. Patients were chronologically divided into training and testing cohorts. High-resolution rectal T₂-weighted imaging (T₂WI) for all enrolled subjects was first reconstructed and optimized using Deep Learning Reconstruction (DLR) technology. On the DLR-reconstructed T₂WI images, the lesion region of interest was manually delineated slice-by-slice using two methods: minimal target delineation (Method 1) and maximal target delineation (Method 2). DL segmentation models were constructed and compared. Subsequently, a DL classification model for preoperative TN staging of rectal cancer was built based on the optimal segmentation method and comprehensively evaluated.

This study included 326 patients, with 143 patients from 2019~2020 as the training set and 183 patients from 2021~2023 as the testing set. The test set showed that the DL segmentation model of Method 2 achieved median DSC, HD95, and ASD of 0.796, 7.730 mm, and 1.420 mm, respectively, which were superior to Method 1 (0.708, 10.931 mm, and 2.388 mm), with higher DSC (P＜0.001) and smaller ASD (P=0.013). The DL classification model was constructed using the segmentation results of Method 2. In the T-staging task, 177 test cases showed that the model achieved an AUC (95%CI) of 0.957 (0.932~0.975), significantly higher than subjective evaluation at 0.811 (0.733~0.882). The diagnostic specificity, sensitivity and accuracy of the model were 91.5%, 95.3% and 93.8%, respectively. In the N-staging task, 182 test cases showed an AUC (95%CI) of 0.837 (0.768~0.897), significantly better than subjective evaluation at 0.625 (0.558~0.803), with diagnostic specificity, sensitivity and accuracy of 98.6%, 63.9% and 91.8%, respectively. DeLong tests showed P＜0.001 for both, and the model demonstrated good calibration and higher clinical net benefit.

DLR-optimized MRI images combined with the DL segmentation model using maximal target delineation can accurately extract quantitative information and significantly improve the diagnostic efficacy of preoperative TN staging of rectal cancer.