Deep Learning For Minimally Invasive Computer Assisted Surgery

Abstract

Detection of surgical instrument has been implemented in minimally invasive computer assisted surgery domain but detection of desired parts of surgical instrument has not been implemented properly. Previous researches have divided surgical instrument into two parts: End-effector and Shaft [12], which are not adequate to detect the components clearly. In this paper, we propose solution to improve accuracy and processing time of instrument detection. The novel detection has been implemented using deep learning algorithms-Convolutional Neural Network (CNN). The CNN uses kernel to perform feature extraction. The feature extraction includes convolution, batch normalisation, ReLu, max pooling and drop. In addition, selective kernel has been used during convolution to detect the parts of surgical instrument. There are four different types of datasets have been used for the execution. The proposed solution has giving promised results as there are nearly 2% improvement in accuracy and nearly 2s drop-in processing time. ReLu activation in convolution network and 20% dropout from output of convolution, not only reduces the processing time but also improved accuracy of detection.