Developing a deep learning-based surgical-skill assessment model focused on instrument handling in laparoscopic colorectal surgery

Introduction

Poor instrument handling, such as "repeatedly makes tentative or awkward moves" and "grasper (frequently) slip," is associated with poor surgical skills. We constructed and applied an automated recognition model of tissue grasping during laparoscopic surgery using computer vision technology to clarify whether automated surgical-skill assessment using the number of tissue grasps could be feasible.

Methods

The number of tissue grasps and classification of success/failure were manually and automatically counted. Intraoperative videos of three groups with obviously different surgical levels (the high-, intermediate-, and low-skill groups) were prepared; an automated distinction between these groups was attempted using the models.

Results

The number of manually counted tissue grasps was significantly higher in the low-skill group than in the other groups, while the number of failed tissue grasps was significantly lower in the high-skill group than in the other groups. The number of automatically counted tissue grasps showed strong correlations with the manually counted ones, whereas the other parameters showed only moderate correlations. The number of automatically counted tissue grasps was significantly higher in the low-skill group than in the other groups, similar to that noted with manual counting. The other automatic parameters showed no results similar to the manual ones.

Conclusion

We successfully constructed automated recognition models of tissue grasping during laparoscopic surgery and found that automated surgical-skill assessment based on the number of tissue grasps could be feasible. However, the results were insufficient for automatically distinguishing between successful/failed tissue grasps. Further improvements in recognition accuracy are required for this model.