Patch-based classification of gallbladder wall vascularity from laparoscopic images using deep learning.

Area of Science:

• Medical Imaging
• Computer Vision
• Surgical Technology

Background:

• Assessing gallbladder (GB) wall vascularity during laparoscopic cholecystectomy (LC) is crucial but challenging due to factors like fatty infiltration or wall thickening.
• Difficulties in visualizing GB wall vessels can impact surgical decisions and outcomes.

Purpose of the Study:

• To develop and evaluate a deep learning approach for automated assessment of gallbladder wall vascularity using intraoperative images from LC.
• To address the limitations in manual visualization of GB wall vessels.

Main Methods:

• Utilized 800 patches and 181 region outlines from the Cholec80 video dataset, annotated by expert surgeons into 2 (low vs. high) and 3 (low, medium, high) vascularity classes.
• Investigated two convolutional neural network (CNN) architectures, incorporating vessel enhancement preprocessing and late fusion post-processing techniques.

Main Results:

• The best CNN model achieved high accuracy: 94.48% for 2-class patch classification and 80.66% for 3-class region classification.
• Model performance was comparable to inter-observer agreement among expert surgeons (91.71% for 2 classes, 78.45% for 3 classes).
• Spatial probability maps were generated, visualizing vascularity class probabilities across the GB wall.

Conclusions:

• This study represents a significant advancement in using computer vision and deep learning for assessing GB wall vascularity from intraoperative LC images.
• The developed deep learning models demonstrate performance on par with expert surgeon agreement.
• The approach holds potential for classifying LC operations and providing context-aware assistance in surgical education and practice.