LAUSR

Pneumothorax is common but a life-threatening thoracic disease, which is difficult to diagnose based on chest X-ray images due to its subtle characteristics and low contrast of the disease regions. We aim to develop a fully automatic pneumothorax segmentation method to assist radiologists for timely and accurate diagnosis of pneumothorax. We propose a two-stage deep learning method. In the first stage, the chest X-ray image is classified as having pneumothorax or not using an ensemble of multiple modified U-Net models. Each model is trained using a multitask learning strategy with two highly correlated tasks: 1) pneumothorax classification and 2) segmentation. The segmentation task helps the model focus on more relevant regions to improve classification accuracy. The second stage performs precise segmentation of pneumothorax regions, which applies another ensemble model that consists of four U-Net-like models and one Deeplabv3+ model. We validated our method by participating in the 2019 SIIM-ACR pneumothorax segmentation challenge. Our method produced a classification area under the receiver operating characteristic curve (AUC) of 0.9795 and a Dice score of 0.8883, which secured the second place among 1475 teams. Fine tuning the models using all available data leads to more accurate results that surpassed all other competing methods. The code is made publicly available online at https://github.com/yelanlan/Pneumothorax-Segmentation-2nd-place-solution. Validation on extra polyp and lung segmentation data further proved the general applicability of the proposed method.