Deep learning-based intraoperative visual guidance model for ureter identification in laparoscopic sigmoidectomy.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Surgery
• Surgical Navigation

Background:

• Accurate identification of the left ureter is crucial during laparoscopic sigmoid resection to prevent surgical injuries.
• Intraoperative ureteral identification remains a challenge in minimally invasive procedures.

Purpose of the Study:

• To evaluate the real-time performance of a deep learning-based computer vision model for left ureter identification.
• To assess the feasibility of using artificial intelligence to enhance surgical safety in laparoscopic sigmoidectomy.

Main Methods:

• A semantic segmentation deep learning model (You Only Look Once versions 8 and 11) was trained on 1237 intraoperative images from 86 laparoscopic sigmoid resection videos.
• Manual annotation by three colorectal surgeons was performed on the surgical video data.
• Model performance was evaluated using per-frame five-fold cross-validation.

Main Results:

• The deep learning model achieved a mean Average Precision (mAP50) of 0.92 and a Dice Coefficient (DC) of 0.90.
• High precision (0.94) and recall (0.88) were recorded, with the highest DC reaching 0.95.
• The model operated at 32 Frames Per Second (FPS), demonstrating real-time capability.

Conclusions:

• Deep learning models, specifically YOLO 8 and 10, can accurately and in real-time identify the left ureter during laparoscopic sigmoidectomy.
• This technology has the potential to assist surgeons with intraoperative image navigation for improved accuracy.
• Limitations include sample size, surgical method diversity, and the need for external validation.