Automatic Surgical Instrument Recognition—A Case of Comparison Study between the Faster R-CNN, Mask R-CNN, and Single-Shot Multi-Box Detectors

Jiann-Der Lee,Chieh-Tsai Wu,Yu-Tsung Hsu,Jong-Chih Chien

doi:10.3390/app11178097

Abstract

In various studies, problems with surgical instruments in the operating room are usually one of the major causes of delays and errors. It would be of great help, in surgery, to quickly and automatically identify and keep count of the surgical instruments in the operating room using only video information. In this study, the recognition rate of fourteen surgical instruments is studied using the Faster R-CNN, Mask R-CNN, and Single Shot Multi-Box Detectors, which are three deep learning networks in recent studies that exhibited near real-time object detection and identification performance. In our experimental studies using screen captures of real surgery video clips for training and testing, this study found that that acceptable accuracy and speed tradeoffs can be achieved by the Mask R-CNN classifier, which exhibited an overall average precision of 98.94% for all the instruments.

Highlights

In various studies on equipment-related incidents in the operating room, it was found that equipment-related type of error occurs in about 15.9% of cases involved in the studies [1,2], mostly due to unavailability of the requested instruments, which caused significant delays in the surgery
Yu et al [10] proposed a modified Single-Show Multi-Box Detector (SSD) (Single-Shot Multi-Box Detector) classifier for real-time processing of videos containing surgical instruments, and it reported that the modified architecture achieved an average precision of 90.08% in a set of images extracted from real surgery video clips
R-CNN, and the Single Shot Multi-Box Detector (SSD) on the detection and identification of surgical instruments used in the Department of Neurosurgery of the Chang Gung

Summary

Introduction

In various studies on equipment-related incidents in the operating room, it was found that equipment-related type of error occurs in about 15.9% of cases involved in the studies [1,2], mostly due to unavailability of the requested instruments, which caused significant delays in the surgery. It would be helpful to assisting nurses as well as the surgeons if a detector can be used to identify and track the surgical equipment before, during, and after the surgery in a near real-time manner using just the videos available during the surgery. This type of detector could help reduce incidents such as accidentally leaving surgical instruments inside of patients. Yu et al [10] proposed a modified SSD (Single-Shot Multi-Box Detector) classifier for real-time processing of videos containing surgical instruments, and it reported that the modified architecture achieved an average precision of 90.08% in a set of images extracted from real surgery video clips. The areas surrounding the objects of interest (called regions of interest, or ROI) are computed from these features

Objectives

Methods

Results

Discussion

Conclusion