Deep learning-based gastrocolic trunk recognition in laparoscopic right hemicolectomy.

BACKGROUND

Intraoperative bleeding from the gastrocolic trunk is a serious complication of laparoscopic right hemicolectomy for right-sided colon cancer. Recognizing the gastrocolic trunk is crucial for surgical safety.

METHODS

We developed and retrospectively evaluated a deep learning model that automatically recognizes the gastrocolic trunk during laparoscopic right hemicolectomy. Still images from laparoscopic right hemicolectomy videos were annotated at each pixel corresponding to the gastrocolic trunk and superior mesenteric vein and input into the deep learning model for segmentation of the gastrocolic trunk. Tasks were categorized as follows based on the segmentation patterns: Task A, gastrocolic trunk and superior mesenteric vein distinguished from the background; Task B, gastrocolic trunk, superior mesenteric vein, and background distinguished individually; and Task C, the gastrocolic trunk distinguished from the superior mesenteric vein and background. Data for this study were obtained from 10 high-volume hospitals in Japan. Images from 43 patients who were diagnosed with right-sided colon cancer and underwent right hemicolectomy between April 2018 and July 2020 were analyzed. Fivefold cross-validation was performed, and the average Dice coefficient, precision, and recall were evaluated.

RESULTS

Overall, 1,625 still images from 43 laparoscopic right hemicolectomy videos were analyzed in this study. The average Dice coefficient for vein segmentations in task A was 0.84, and those for gastrocolic trunk segmentations in tasks B and C were 0.61 and 0.58, respectively.

CONCLUSIONS

We developed a deep learning model that can identify and visualize the gastrocolic trunk in surgical videos of laparoscopic right hemicolectomy, which can improve procedural safety. Future studies should confirm whether this model can effectively reduce the risk of accidental intraoperative bleeding during laparoscopic right hemicolectomy in real-world clinical settings.