Artificial intelligence real-time overall recognition model for total mesorectal excision

Total mesorectal excision (TME), first proposed by Heald in 1982, remains the standard surgical approach for rectal cancer. However, accurately and quickly identifying and separating the correct planes while avoiding damage to blood vessels, nerves, and ureters remains challenging in TME. This study aimed to develop an artificial intelligence recognition model for the different structures involved and to evaluate its performance through internal and external verifications.

A database of pelvic autonomic nerves, arterial and venous identification, ureters, the Toldt fascia, and separation layer during TME surgery was established by retrospectively collecting intraoperative images and videos of patients with rectal cancer from January 2016 to April 2024. An overall identification and navigation model for TME surgery was established through steps such as image sketching, model training and evaluation, and clinical application.

A total of 6700 high-quality intraoperative images from 325 patients in the training group were collected for this study. After 1000 iterations, the current model was obtained. The parameters for the model, namely mean intersection over union, recall, precision, and F1, were as follows: ureter (0.7086, 0.7601, 0.9128, and 0.8295), artery (0.8095, 0.8789, 0.9111, and 0.8947), vein (0.8114, 0.8599, 0.9350, and 0.8959), Toldt fascia (0.8985, 0.9532, 0.9400, and 0.9466), and separation layer (0.7281, 0.8191, 0.8676, and 0.8427).

This study successfully established a comprehensive recognition model for TME surgery, in which different structures were recognized as different colors. The successful establishment of this model is expected to prevent intraoperative structural damage and promote TME surgery standardization.