Empirical Assessment of Transfer Learning Techniques for Surgical Tools Classification

Abstract

Automated surgical tool classification in the medical images is a real-time computerized assistance for the surgeons in performing different operations. Deep learning has evolved in every facet of life due to availability of large datasets and emergence of Convolutional Neural Networks (CNN) that have paved the way for development of different image related processes. In the medical field there are number of challenges such as non-availability of datasets, image annotation requires extensive time, imbalanced data. Transfer learning is the process of applying existing pretrained models to the new problem. It is useful in those scenarios where the large datasets are not available, or the new dataset shares visual features with the existing dataset on which the model is pretrained. Most of the pretrained models are trained on ImageNet which is a largescale dataset (1.2 million labelled training images). In this paper we evaluated and explored two different CNN architectures namely VGG16 and MobileNet-v1-1.0-224 on subset of surgical toolset. This paper presents comparative analysis of the techniques using learning curves and different performance metrics.