Deep Learning-Based Real-Time Ureter Identification in Laparoscopic Colorectal Surgery.

Area of Science:

• Medical imaging
• Artificial intelligence in surgery
• Surgical navigation

Background:

• Iatrogenic ureteral injury is a significant risk during abdominopelvic surgery.
• Accurate intraoperative identification of ureters is crucial for preventing such injuries.
• A novel model was developed to mitigate this surgical complication.

Purpose of the Study:

• To evaluate the efficacy of a deep learning model, UreterNet, for ureter recognition in laparoscopic colorectal surgery videos.
• To determine if UreterNet can accurately identify ureters in real-time surgical video feeds.

Main Methods:

• A deep learning-based semantic segmentation algorithm, UreterNet, was developed using a convolutional neural network architecture (Feature Pyramid Networks).
• The model was trained and validated on 14,069 annotated images from 304 surgical videos.
• Performance was assessed using precision, recall, and Dice coefficient, with real-time inference speed measured.

Main Results:

• UreterNet achieved a precision of 0.712, recall of 0.722, and Dice coefficient of 0.716 on the test dataset.
• The model demonstrated real-time performance, inferring ureter pixels in 71 milliseconds and displaying results in 143 milliseconds.
• These results indicate successful ureter segmentation in laparoscopic colorectal surgery videos.

Conclusions:

• UreterNet offers a noninvasive method for intraoperative ureter identification during laparoscopic colorectal surgery.
• The model has the potential to enhance surgical safety by reducing the risk of iatrogenic ureteral injury.
• Further clinical validation is required to confirm UreterNet's impact on reducing actual iatrogenic ureteral injury rates.