Motion Primitive Learning for Robot-Assisted Surgery

Abstract

Although Minimally Invasive Robotic Surgery (MIRS) had been researched for more than 25 years, and commercial and advanced research systems exist, Artificial Intelligence and Big Data approaches open up new fields of research. For instance, the collection of Big Data from procedures aid the new field of research on surgical activity recognition (SAR). SAR could be the basis to provide assistance functions to the operating surgeon to increase performance. SAR is in essence a problem of recognizing repeatedly incoming sequential data. This has been previously approached using Hidden Markov Models (HMMs). In this work, a motion recognition approach, called Incremental Motion Primitive Learning (MPL), and two Linear Discriminant Analysis (LDA) based methods are adapted and implemented to aid SAR. These methods are compared to methods proposed in the literature, by evaluating them on a publicly available dataset (JIGSAWS). The LDA based approaches show promising results up to 94% accuracy, while the MPL approaches perform not as well. Shortcomings and possible improvements are discussed and future work on the methods have been identified. Additionally, a new surgical activity dataset using the DLR MiroSurge System is collected. This dataset augments the commonly used kinematic data with dynamic, contextual and low-level hardware data. This dataset can be a basis to support the research on SAR, e.g. on feature selection.