Machine learning models that detect surgical activities in endoscopic videos are instrumental in scaling post-surgical video review tools that help surgeons improve their practice. However, it is unknown how well these models generalize across various surgical techniques practiced at different institutions. In this paper, we examined the possibility of using surgical site information for a more tailored, better-performing model on surgical procedure segmentation. Specifically, we developed an ensemble model consisting of site-specific models, meaning each individual model was trained on videos from a specific surgical site. We showed that the site-specific ensemble model consistently outperforms the state-of-the-art site-agnostic model. Furthermore, by examining the representation of video-frames in the latent space, we corroborated our findings with similarity metrics comparing videos within and across sites. Lastly, we proposed model deployment strategies to manage the introduction of videos from a new site or sites with insufficient data.
|