Fusing CNN features to improve generalisability for surgical tool classification

Abstract

Surgical tool classification is a fundamental component for understanding surgical workflow of laparoscopic interventions. Image-based approaches using convolutional neural networks (CNN) have been prominent with availability of computing infrastructures and achieved high performance. However, such approaches need to be assessed in terms of robustness and generalisability to new data sources. Previous works have revealed low generalisation performance of CNN base models. This work proposes a method to enhance CNN generalisability by fusing features from multiple intermediate layers. Experimental results showed good improvement in generalisation performance on data obtained from new clinics and unseen types of procedures.