Pneumothorax Segmentation from Chest X-Rays Using U-Net/U-Net++ Architectures

Abstract

Image segmentation is a vital part of many visual understanding systems. It has several advantages such as easy data storing, faster processing, and a multiple copy system. A large number of image segmentation methods can be seen in the advanced medical field. Segmentation of medical images requires extreme precision and error-prone results. Also, semantic segmentation has shown the most precise results in these terms. In this study, we assessed chest x-ray images for lung (Pneumothorax) semantic segmentation. For this case, we employed U-Net++ with an aim to segmentation the x-ray images for detecting Pneumothorax and also to identify their positions in the human body. Additionally, we incorporated several image recognitions models Res-Net34, Xception, and Inception V3 within U-Net++ architecture and examined the model which delivers the least amount of loss with enhanced accuracy. The outcomes of the study will be favorable for clinicians aimed at accurate diagnosis but also to reduce diagnostic limitations.