Deep Learning-Based Prediction System for Surgical Difficulty in Rectal Cancer Patients Using MRI Pelvimetry

This study presents a system that automatically predicts the difficulty of laparoscopic total mesorectal excision (TME) using magnetic resonance imaging (MRI) pelvimetry and the clinical characteristics of the patients with rectal cancer, using deep learning (DL) technology and statistical analysis.

Colorectal MRI data were collected from patients with rectal cancer and input into to the segmentation model for DL training, followed by automatic measurement of pelvimetry parameters from the segmentation results. The measured parameters and preoperative clinical data of the patients were statistically analyzed to identify clinically correlated predictors of surgical difficulty. Finally, using the selected predictors, a logistic nomogram was generated for the prediction of surgical difficulty.

The segmentation models yielded Dice similarity coefficients (DSC) of 92.7%-95.7%, and the automatic measurement results exhibited high correlation with the manual measurement results. The identified predictors were age, tumor location from the anal verge, interspinous distance, angle Δ, and mesorectal fat area. The C-index of the generated nomogram was 0.852.

The system provides standard guidance for the examination of the difficulty of TME that could potentially be used in clinical settings. A follow-up study could develop this idea to fully automate the entire process with a simple user interface to provide objective predictions and explanations.