Zhonghua Wei Chang Wai Ke Za Zhi. 2026 Jan 25;29(1):76-82. doi: 10.3760/cma.j.cn441530-20251015-00382.
ABSTRACT
Objective: This study processes and analyzes rectal MRI images of patients with mid-to-low rectal cancer using deep learning technology, and integrates these data with clinical baseline information to construct a fully automated end-to-end prediction model. The model is designed to assist colorectal surgeons in preoperatively assessing surgical difficulty and selecting the optimal surgical approach. Methods: We prospectively collected data from patients with mid-to-low rectal cancer who underwent laparoscopic total mesorectal excision (TME) and had been graded according to the surgical difficulty system recorded in Division of Colorectal Surgery, Department of General Surgery, Peking Union Medical College Hospital, between March, 2019 and May, 2025. Inclusion criteria: (1) age 18-75 years; (2) tumor lower edge within 10 cm of the anal verge as measured by rectal MRI; (3) pathologically confirmed rectal adenocarcinoma; (4) complete, accessible preoperative rectal MRI DICOM images; and (5) tumor invasion depth of T1-4aNanyM0. Exclusion criteria: (1) synchronous or metachronous multiple primary colorectal cancer with concurrent surgery; (2) Any surgery other than TME; (3) tumor involvement of surrounding organs requiring combined organ resection; (4) unfitness for laparoscopic surgery (e.g. extensive adhesions from previous abdominal surgery, contraindications to pneumoperitoneum for various reasons, etc.); and (5) Robot-assisted radical resection of rectal cancer. Included patients were divided into training and test datasets, and deep learning techniques (rectal MRI image annotation, image preprocessing, data augmentation, and feature extraction) were used for model construction. Results: A total of 366 patients were included, with 253 males. The median BMI was 24.1 (22.0, 26.6) kg/m², and the median distance from the tumor lower edge to the anal verge was 6.5 (4.7, 7.8) cm. A total of 288 patients received neoadjuvant chemoradiotherapy. Based on intraoperative difficulty grade, patients were divided into the difficult group (199 cases) and the nondifficult group (167 cases). Compared to the nondifficult group, the difficult group showed several statistically significant differences (all P<0.05): higher proportion of males [86.9%(173/199) vs. 47.9%(80/167), χ²=64.813, P<0.001]; higher BMI [25.4 (23.2, 27.6) kg/m² vs. 23.1 (21.2, 25.2) kg/m², Z=-6.082, P<0.001]; and higher proportion of neoadjuvant chemoradiotherapy [88.9% (177/199) vs. 66.5%(111/167), χ²=27.357, P<0.001]. However, there was no statistically significant differences in the distance from the tumor lower edge to the anal verge between the two groups [6.4 (4.7, 7.9) cm vs. 6.6 (4.7, 7.7) cm, Z=-0.001, P=0.999]. Importantly, our surgical difficulty prediction model achieved an accuracy of 0.729, a precision of 0.684, a specificity of 0.521, a recall of 0.915, an F1-score of 0.782, and an AUC of 0.83. Conclusions: We proposed a prediction model with reasonable accuracy using artificial intelligence that can assist surgeons in determining surgical difficulty and choosing the optimal surgery approach for mid-low rectal cancer.
PMID:41566184 | DOI:10.3760/cma.j.cn441530-20251015-00382
