Algorithms in Surgical Action Recognition: A Survey

Abstract

Machine learning and medical image processing technology continue to promote the advancement of scene perception method. Among the field of scene perception, surgical action recognition is finer-grained and conductive to surgical guidance and evaluation, promoting the realization of automatic surgical robot. In this work, an overall review of surgical action recognition is provided, including the methods, datasets, metrics, and prospects. From the perspective of labeling degree of datasets, the methods are categorized: supervised learning, weakly-supervised or unsupervised learning, and reinforcement learning. Correspondingly, their algorithms, performances and contributions are concluded. Then, the commonly used datasets and metrics in surgical action recognition are summarized. Finally, the prospect is discussed. This survey highlights the relevant corner stones and latest researches in the field, covers the widely used datasets and metrics, offers valuable insight into the prospects, and aims to gain more attention of researchers to surgical action recognition.