The Effect of Background Pattern on Training a Deep Convolutional Neural Network for Surgical Tool Detection

Abstract

Surgical tool presence detection in laparoscopic videos plays a key role in developing context-aware systems (CASs). These systems are designed to support surgical team inside operating rooms and increase the efficiency of surgical workflow. Convolutional neural networks (CNNs) have shown robust performance in detecting surgical tools in laparoscopic images. However, imbalanced data sets are still influencing the training process of the CNN models. In this work, data augmentation methods based on generated artificial images as training patterns by substituting the image background by uniform, random or original background patterns are investigated. First experimental results show different effects on the training process. Easily, an improvement of 10% in classification accuracy could be achieved when the network was trained on augmented data.