Pneumothorax Segmentation In Chest X-Rays Using UNet++ And EfficientNet

Abstract

Pneumothorax or collapsed lung, is a condition that occurs when air enters the space between the chest wall and the lung. Normally, looking at a chest X-ray image is the best way for an expert or experienced radiologist to make sure that one has a collapsed lung or not. However, in certain cases, it is difficult for the experts to diagnose pneumothorax since other medical conditions may look similar. Moreover, diagnosing quickly this disease is a hard problem in the underdeveloped regions because of the lack of the experienced radiologists. Lately, with the growth of large neural network architectures and medical imaging datasets, deep learning has been providing diagnostic support systems in detecting and locating pneumothorax with high accuracy. In this paper, an image segmentation model was proposed to support doctors in taking crucial decision by determining pneumothorax on a chest X-ray image. The UNet++ architecture has been used with EfficientNet (EfficientNet-B4) as a backbone which is pre-trained on ImageNet dataset. The chest X-ray dataset of 2019 SIIM-ACR Pneumothorax Segmentation Challenge, which contains 12047 training images and 3205 testing images, was used for testing. This method achieves 0.8544 mean Dice coefficient placing it among the top 1,7% of competitors with a rank of 26 out of 1475 teams.