Artificial Intelligence-Guided Total Mesorectal Excision: Development of a Deep Learning Model to Identify the Pelvic Fascial Plane Anatomy

Yuta Suzuki; Daichi Kitaguchi; Norihito Kosugi; Hiro Hasegawa; Nobuyoshi Takeshita; Yusuke Kinugasa; Masaaki Ito

doi:10.1002/ags3.70064

Artificial Intelligence-Guided Total Mesorectal Excision: Development of a Deep Learning Model to Identify the Pelvic Fascial Plane Anatomy

Ann Gastroenterol Surg. 2025 Jul 9;9(6):1225-1232. doi: 10.1002/ags3.70064. eCollection 2025 Nov.

Authors

Yuta Suzuki^{1

2}, Daichi Kitaguchi^{1

3}, Norihito Kosugi¹, Hiro Hasegawa^{1

3}, Nobuyoshi Takeshita¹, Yusuke Kinugasa², Masaaki Ito^{1

3}

Affiliations

¹ Department for the Promotion of Medical Device Innovation National Cancer Center Hospital East Chiba Japan.
² Department of Gastrointestinal Surgery, Graduate School of Medical and Dental Sciences Institute of Science Tokyo Tokyo Japan.
³ Department of Colorectal Surgery National Cancer Center Hospital East Chiba Japan.

Abstract

Aim: Total mesorectal excision entails dissection of the "holy plane" without damaging the mesorectal and parietal pelvic fasciae. We aimed to train a deep learning model to identify the anatomical landmarks of the pelvic fascial surface during total mesorectal excision and evaluate its performance to validate artificial intelligence-guided total mesorectal excision.

Methods: In this single-center, retrospective, observational study, a deep learning model that can automatically identify the anatomical landmarks in the pelvic fascial plane during laparoscopic total mesorectal excision in real time was developed. Target anatomical landmarks included the mesorectal fascia, parietal pelvic fascia, and the holy plane between them. Feature Pyramid Networks and EfficientNetB7 were adopted as the neural network architecture and backbone network, respectively. The surgical videos used for training and validation data were different. The Dice Similarity Coefficient, recall, and Normalized Surface Dice, which were calculated using five-fold cross-validation, were used as the evaluation metrics.

Results: Overall, 2861 images from 157 surgical videos were annotated and used for the training and validation datasets. In the semantic segmentation task, the Dice Similarity Coefficients for the mesorectal fascia, parietal pelvic fascia, and holy plane were 90.4%, 90.6%, and 68.5%, respectively. The Normalized Surface Dices for the three fasciae were 70.4%, 73.2%, and 54.6%, respectively.

Conclusions: We constructed a deep learning model that recognizes three regions for total mesorectal excision-the mesorectal fascia, parietal pelvic fascia, and intervening holy plane-with a relatively high Dice Similarity Coefficient, successfully providing proof of concept for artificial intelligence-guided total mesorectal excision.

Keywords: artificial intelligence; coloproctectomy; deep learning; rectal neoplasms; total mesorectal excision.