Evaluation of deep learning techniques on a novel hierarchical surgical tool dataset

Abstract

A new hierarchically organised dataset for artificial intelligence and machine learning research is presented, focusing on intelligent management of surgical tools. In addition to 360 surgical tool classes, we create a four level hierarchical structure for our dataset defined by 2 specialities, 12 packs and 35 sets. We employ different convolutional neural network training strategies to evaluate image classification and retrieval performance on this dataset, including the utilisation of prior information in the form of a taxonomic hierarchy tree structure. We evaluate the effects of image size and the number of images per class on model predictive performance. Experiments with the mapping of image features and class embeddings in semantic space using measures of semantic similarity between classes show that providing prior information results in a significant improvement in image retrieval performance on our dataset.