Efficient segmentation of intraoperative anatomical landmarks in laparoscopic cholecystectomy based on deep learning.

BACKGROUND: Laparoscopic cholecystectomy is the gold standard procedure for the treatment of benign gallbladder diseases, but there is the risk of intraoperative bile duct injury, which can lead to surgical failure and cause significant social and economic burden. When surgeons rely on visual inspection to identify tissue structures during laparoscopic cholecystectomy, subjective factors such as experience, psychological factors, and fatigue can compromise the intraoperative recognition of anatomic landmarks. The positioning of anatomical landmarks by the surgeon in the pre-dissection phase of laparoscopic cholecystectomy is relatively vague and requires step-by-step exploration as the surgery progresses, becoming clearer in the post-dissection phase.

METHODS: To alleviate the pressure on surgeons during procedures, this study aimed to achieve real-time intraoperative navigation during laparoscopic cholecystectomy by dynamically identifying and annotating key anatomical landmarks, including the gallbladder, Calot's triangle, and common bile duct. The study proposed a novel semantic segmentation neural network called the Channel Attention Pyramid Scene Parsing Plus Network. The network utilized pooling layers with different scales and assigned non-equal weights to extract feature information. Additionally, a spatial channel attention module was added to accurately capture important features or contextual information, improving the model's performance and effectiveness. Training was conducted using video frames from the pre-dissection phase, while testing used video frames from the post-dissection phase.

RESULTS: All models were subjected to a 10-fold cross-validation on 1425 selected frames from 132 laparoscopic cholecystectomy videos, with training and validation conducted in two separate laparoscopic cholecystectomy stages. The proposed model CPPN achieved a mean intersection over union of 0.855 (±0.03), outperforming other segmentation neural networks. The model demonstrated optimal performance across most metrics, with an intersection over union of 0.881 (±0.01) for the gallbladder, 0.769 (±0.03) for Calot's triangle, and 0.813 (±0.02) for the common bile duct.

CONCLUSION: The intelligent segmentation algorithm proposed in this study has achieved the highest mean intersection over union, surpassing other models. It shows promise in assisting surgeons with the real-time assessment of critical anatomical landmarks within Calot's triangle. This advancement could potentially reduce the risk of common bile duct injury by facilitating a more intuitive dissection of Calot's triangle. Furthermore, it aids in the visual inspection during laparoscopic cholecystectomy procedures.