Pneumothorax Detection and Localization in Chest Radiographs: A Comparison of Deep Learning Approaches

André Gooßen, Hrishikesh Deshpande, Tim Harder, Evan Schwab, Ivo Baltruschat, Thusitha Mabotuwana, Nathan Cross, Axel Saalbach

Keywords: Deep Learning, Convolutional Neural Networks, Fully Convolutional Networks, Multiple-Instance Learning, ResNet, U-Net, Pneumothorax, Chest X-Ray

TL;DR: We evaluate three deep learning architectures for the detection and localization of pneumothorax in chest X-ray images.

Abstract: Pneumothorax is a critical condition that requires timely communication and immediate action in order to prevent significant morbidity or patient death. Thus, early detection is crucial. For the task of pneumothorax detection, we measured the performance of three different deep learning techniques: convolutional neural networks, fully convolutional networks, and multiple-instance learning. In a five-fold cross-validation with ROC analysis on a dataset consisting of 1003 chest X-Ray images we measured AUCs of 0.96, 0.92, and 0.93 for the three methods, respectively. We discuss the classification as well as localization performance of these approaches as well as an ensemble of aforementioned techniques.

Code Of Conduct: I have read and accept the code of conduct.

Remove If Rejected: Remove submission from public view if paper is rejected.