Abnormality Detection and Localization in Chest X-Rays using Deep Convolutional Neural Networks

Chest X-Rays (CXRs) are widely used for diagnosing abnormalities in the heart lung area. Automatically detecting these abnormalities with high accuracy greatly enhance real world diagnosis processes. Lack of standard publicly dataset and benchmark studies, however, makes it difficult to compare detection methods. In order to overcome these difficulties, we have a publicly available Indiana CXR, JSRT and Shenzhen dataset and studied performance of known deep convolutional network (DCN) architectures on abnormalities. We find that the same DCN architecture doesn't perform across all abnormalities. Shallow features or earlier layers consistently higher detection accuracy compared to deep features. We have also found models to improve classification significantly compared to single. Combining these insight, we report the highest accuracy on chest X-Ray detection on these datasets. We find that for cardiomegaly, the deep learning method improves the accuracy by a staggering 17 point compared to rule based methods. We applied the techniques to problem of tuberculosis detection on a different dataset and achieved the accuracy. Our localization experiments using these trained classifiers that for spatially spread out abnormalities like cardiomegaly and edema, the network can localize the abnormalities successfully most the time. One remarkable result of the cardiomegaly localization is that the and its surrounding region is most responsible for cardiomegaly, in contrast to the rule based models where the ratio of heart and area is used as the measure. We believe that through deep learning based and localization, we will discover many more interesting in medical image diagnosis that are not considered traditionally.