Multi-Class Detection of Laparoscopic Instruments for the Intelligent Box-Trainer System Using Faster R-CNN Architecture

Abstract

Laparoscopic Surgical Box-Trainer devices have been used by surgery residents to learn specific skills not traditionally taught to surgeons. Assessment of performance, however, is crude, frequently focusing on speed alone or subjective observations. For a better, objective assessment, the residents' efficiency should be recorded and have the process be tracked and have a system in place to provide consistent automated assessment and analysis. In this paper, we propose a novel framework for the detection and recognition of multi-class laparoscopic instruments for our Intelligent Box-Trainer System. This framework is based upon the Faster R-CNN architecture and RESNet-50 for an open-source module with our custom dataset (AR-Set). Despite a relatively limited number of training examples, experimental results have proved that our approach is effective for locating regions of interest and detecting multi-class instruments. This research is a cooperation between the Department of Electrical and Computer Engineering and the Department of Surgery of the Homer Stryker M.D. School of Medicine, at WMU.