CONSTRAINED CONTRASTIVE REPRESENTATION: CLASSIFICATION ON CHEST X-RAYS WITH LIMITED DATA

One of the challenges in the field of medical image classification is the expensiveness of labeled data. Most of the previous computer-aided diagnostic methods are based on a paradigm of object detection. Such ways need tons of labeled sample images with positioning annotations, which always need practicing radiologists to process data manually. We focus on Chest X-ray(CXR) images classification and propose an effective framework for lung disease diagnosis based on a self-supervised feature extracting mechanism trained in a constrained contrastive method. Our proposed framework can train on a relatively small dataset in a semi-supervised way and without any positioning annotation. We experiment with the proposed framework on several lung disease diagnosis tasks, including pneumonia and tuberculosis diagnosis, and obtain state-of-the-art results even outperform previous supervised transfer-learning methods.