Video-based skill acquisition assessment in laparoscopic surgery using deep learning

Purpose Surgical skill acquisition is based on the repetitive execution of a surgical task. Each trainee starts with a different skill level and has a unique pace in attaining the necessary motor skills. Current training curriculums, however, lack adaptability to individual learning abilities. They are also time- and resource-intensive for trainees and expert surgeons. The purpose of this study is to provide an individualized and automated assessment of the surgical skills from the surgical learning curves of the subjects using deep learning and video data feeds. Methods The proposed framework consists of two steps. First, given that surgical training is predominantly on surgical simulators, our pipeline was tested on the videos of a laparoscopic suturing simulation under the context of the Fundamentals of Laparoscopic Surgery (FLS). Twenty-three subjects performed FLS suturing with the intracorporeal knot-tying task over two weeks for more than 1600 trials. A self-supervised contrastive model extracted spatiotemporal features from raw videos of suturing trials. Second, a 1D residual neural network was used to utilize these features to classify surgical procedure outcomes and directly predict performance scores. Results Our findings showed that videos of surgical skill training could sufficiently be used to classify the Pass and Fail outcomes of the procedure with an Accuracy of 0.878 ± 0.002. Furthermore, our network can predict the performance scores of the trainees with a Spearman Correlation Coefficient of 0.746 ± 0.002. This approach offers an objective, automated strategy for efficient surgical skill acquisition prediction. Conclusion We show that deep learning techniques and video alone can be used to predict surgical skill training task outcomes accurately and reliably. From here, we seek to enable customization of the training program.