EP541 - Establishment and Validation of an Artificial Intelligence Prediction Model for Surgical Difficulty in Mid-Low Rectal Cancer: A Single-Center Cohort Study

ePoster Presenter(s)

• 

  Zhen Sun

  Residient
  Peking Union Medical College Hospital
  Beijing, Beijing, China (People's Republic)

  Disclosure information not submitted.

  View ePoster

Author(s)

• BS

  Bo Sun, n/a

  artificial intelligence
  peking union medical college hospital
  Beijing, Beijing, China (People's Republic)

  Disclosure information not submitted.
• KL

  Kexuan Li, n/a

  Resident
  Peking Union Medical College Hospital
  Beijing, Beijing, China (People's Republic)

  Disclosure information not submitted.
• GL

  Guole Lin, Professer

  Chief Physician
  Peking Union Medical College Hospital
  Beijing, Beijing, China (People's Republic)

  Disclosure information not submitted.
• BW

  Bin Wu, n/a

  senior
  Peking Union Medical College Hospital
  Beijing, Beijing, China (People's Republic)

  Disclosure information not submitted.
• LX

  Lai Xu, Colorectal specialist

  No. 1, Shuaifuyuan, Dongcheng District, Beijing 100730, China
  Peking Union Medical College Hospital
  Beijing, Beijing, China (People's Republic)

  Disclosure information not submitted.
• JL

  Junyang Lu, n/a

  Vice professor
  Peking union medical college hospital
  Beijing, Beijing, China (People's Republic)

  Disclosure information not submitted.
• GZ

  Guannan Zhang, n/a

  Attending
  Pumch
  bwijing, Beijing, China (People's Republic)

  Disclosure information not submitted.
• WL

  Weimin Liu, n/a

  PhD Student
  Beihang University
  Beijing, Beijing, China (People's Republic)

  Disclosure information not submitted.

Senior Author(s)

• YX

  Yi Xiao, n/a

  senior
  Peking Union Medical College Hospital
  Beijing, Beijing, China (People's Republic)

  Disclosure information not submitted.

Purpose/Background: The standard surgical treatment for mid-low rectal cancer is total mesorectal excision (TME). Due to location of tumor, the anatomical confine of pelvic and thick mesorectal fat, the difficulty of operation is significantly increased in the TME for patients with mid-low rectal cancer. There has been an increasing interest in factors affecting the difficulty of performing surgery in the pelvic cavity. Artificial intelligence has enabled rapid advances in surgery, but its role in assessing the difficulty of laparoscopic rectal surgery is unclear. This study aims to utilize deep learning technology for intelligent processing and analysis of rectal MRI images from patients with mid-low rectal cancer, and to deeply integrate clinical baseline data to construct a fully automated predictive model for the entire process.

Methods/Interventions: Patients who underwent laparoscopic rectal surgery from March 2019 to May 2025 were included and divided into non-difficult group and difficult group according our surgical difficulty grading system (Grade I represented no surgical difficulty (non-difficult group), whereas grade II-IV represented patients with surgical difficulty (difficult group).). Artificial intelligence was used to build a prediction model of difficult pelvis, and the ability of the model to stratify the difficulty of surgery is evaluated.

Results/Outcomes: The 366 enrolled patients were divided into difficult group (n=167) and non-difficult group (n=199). One hundred and seventy-three of the patients in the difficult group (86.9%) were male. The median BMI was 25.4 (23.2-27.6) kg/m2 and the median distance of the lower edge of the tumor from the anal verge was 6.4 (4.7-7.9) cm. A total of 177 patients (88.9%) received neoadjuvant therapy. In the non-difficult group, a total of 80 patients (47.9%) were male, the median BMI was 23.1 (21.2-25.2) kg/m2 and the median distance of the lower edge of the tumor from the anal verge was 6.6 (4.7-7.7) cm. A total of 111 patients (66.5%) received neoadjuvant therapy. The two patient groups showed statistically significant differences in the proportion of male patients, the rate of neoadjuvant chemoradiotherapy, and the median BMI. After training and testing, the accuracy of the model was 0.729, the precision was 0.684, the specificity was 0.521, the recall was 0.915, the F1-score was 0.782, the area under the receiver operating curve was 0.83 and the average precision was 0.82.

Conclusion/Discussion: We successfully proposed a prediction model with reasonable accuracy using artificial intelligence, which can assist surgeons determine surgical difficulty and choose the optimal surgery approach for mid-low rectal cancer patients.

The Performance of Model
The ROC curve (a), P-R curve (b) and Confusion matrix (c) of the merged model