Surgical Suturing Skill Assessment Using Estimated Hand Roll Angle from a Deep-learning Computer Vision Algorithm

This paper proposes a deep-learning computer vision algorithm to estimate hand roll angles for metric-based assessment of surgical suturing skills. The number of rolls metric, previously calculated directly from IMU data, counts the number of hand roll reversals during a single suture. To calculate this metric using computer vision, we apply a deep-learning algorithm that can reliably estimate hand roll angles after training on suturing videos collected on the SutureCoach simulator. Results show that the estimation accuracy of the deep-learning algorithm is robust to different video backgrounds. The number of rolls metrics were used to analyze suturing performance in the SutureCoach dataset, which includes attending and resident surgeons and novices. The means of number of rolls based on computer vision differs between most skill levels at both surface and depth conditions, a pattern which holds for number of rolls based on the IMU as well. The proposed algorithm provides a solution for non-contact hand roll angle estimation, which opens up the possibility of inter-operative surgical skill assessment. The code is available at https://github.com/axin233/hand_roll_estimation