Pneumothorax segmentation of chest X-rays using improved UNet++

Abstract

Pneumothorax is a life-threatening medical emergency that results in a state of pneumoperitoneum in the chest due to the entry of gas into the pleural cavity. Pneumothorax usually overlaps with tissues such as ribs and clavicles, which are usually difficult to identify on chest X-rays and have a large clinical underdiagnosis. In recent years, breakthroughs have been achieved in many medical image segmentation tasks using deep learning methods. However, the blurred boundary and tissue overlap in chest X-ray pneumothorax segmentation make it difficult for many algorithms to achieve better results in pneumothorax segmentation. To address these problems, we propose a deep learning network ResNeSt-UNet++ based on UNet++ and ResNeSt. In detail, ResNeSt-UNet++ designs a context-aware feature encoder with residual blocks to extract multi-scale features and introduces hybrid jump paths to obtain and fuse image features at different scales. Moreover, the ResNeSt- UNet++ network uses spatial and channel squeezes and excitation (scSE) modules as decoders to refine. Further, ResNeSt-UNet++ defines a loss function based on Binary Cross Entropy to train the network. The Dice similarity coefficient values and IOUs obtained using this method on the X-ray pneumothorax dataset are 88.31% and 83.1%, respectively, which achieve better performance in pneumothorax segmentation compared with networks such as traditional FPN and UNet++. The experimental results show that this method can segment X-ray pneumothorax with high accuracy and help doctors to provide a reference for accurate judgment of X-ray pneumothorax.