Assessing Generalisation Capabilities of CNN Models for Surgical Tool Classification

Abstract

Accurate recognition of surgical tools is a crucial component in the development of robust, context-aware systems. Recently, deep learning methods have been increasingly adopted to analyse laparoscopic videos. Existing work mainly leverages the ability of convolutional neural networks (CNNs) to model visual information of laparoscopic images. However, the performance was evaluated only on data belonging to the same dataset used for training. A more comprehensive evaluation of CNN performance on data from other datasets can provide a more rigorous assessment of the approaches. In this work, we investigate the generalisation capability of different CNN architectures to classify surgical tools in laparoscopic images recorded at different institutions. This research highlights the need to determine the effect of using data from different surgical sites on CNN generalisability. Experimental results imply that training a CNN model using data from multiple sites improves generalisability to new surgical locations.

Keywords

• cnn generalisability
• surgical tool detection
• laparoscopic images.