A deep learning based instrument detection approach for automated surgical systems

Abstract

The demand of automated systems based on artificial intelligence in healthcare has remarkably increased in the last few years. Due to the growing shortage of skilled workers and the associated error potentials, the reduction of the workload is essential for the care of patients. Surgery assisting tasks could be automated in order to overcame these negative effects. This study presents the development and evaluation of a deep learning system for the recognition of surgical instruments. Already implemented algorithms based on Convolutional Neural Networks (CNN) were used. The object detection was carried out with YOLOv5. Altogether 18 models have been trained on a self-generated dataset of around 800 images. A mean average precision (mAP) of 0.978 for the recognition of three classes, and an mAP of 0.874 for the recognition of six classes was achieved.