Semantic segmentation networks of organs in minimally invasive surgery

Abstract

Background: Minimally invasive surgery (MIS) and robot-assisted surgery have gained recognition as procedures safer than traditional laparotomy which facilitate faster patient recovery.However, MIS limits the sense of the surgeon.Therefore, a computer-assisted algorithm is proposed to assist in this surgery.With the advent of convolutional neural networks, machine vision technology has become an attractive option.Materials and Methods: We use four networks, TernausNet, TernausResNet, LinkNet, and DeepLab V3+, to predict organ segments in endoscopy images.Furthermore, endoscopy images have several issues such as noise, hemorrhage, and shading.Therefore, we perform preprocessing and draw parallels between the images with and without preprocessing. Result:The network with the lowest performance is TernausNet; the performances of the other three networks show marginal differences.The most significant factor for predicting performance is the encoder network.All networks demonstrate reliable performance with a minimum intersection over union score of 0.68 in TernausNet. Conclusion:The segmentation of organs in images can be used for the quantitative evaluation of surgery and to help surgeons understand anatomy.

Ähnliche Arbeiten