Inter-hospital transferability of AI: A case study on phase recognition in cholecystectomy

Identifying surgical phases is a crucial component of surgical workflow analysis, facilitating the automated evaluation of surgical procedures' performance and efficiency. A significant challenge in developing neural networks for surgical phase recognition lies in the scarcity of training data and the large variation in surgical techniques among surgeons. Consequently, it is imperative for these networks to possess generalization capabilities across diverse datasets. In this paper, we analyze the transferability of trained phase recognition models, using cholecystectomy as a case study.

We employed datasets comprising 104 publicly available surgeries from three different centers for training and conducted multiple experiments using 21 videos of surgeries we recorded ourselves for evaluation. A two-stage deep learning architecture was employed, using a ResNet50 backbone followed by a multi-stage Temporal Convolutional Network (MS-TCN). Several experiments were conducted, including training solely on MHB data, training exclusively on public data, and training on a combination of both with an additional fine-tuning approach.

Models trained solely on MHB data achieved an accuracy of approximately 79.7%, while those trained on public data alone performed significantly worse when applied to MHB data. The best performance was obtained by retraining on a combined dataset. The results indicate that it is possible to transfer models to new environments (operating rooms or clinics) and surgeons by using public data, and incorporating site-specific data improves model transferability.

The results demonstrate that leveraging diverse training data, including institution-specific videos, is crucial to develop robust and transferable AI models for surgical phase recognition, thereby enhancing the potential of automated decision-support systems across different clinical environments.