Abstract
Chest X-ray (CXR) is the most typical medical image worldwide to examinevarious thoracic diseases. Automatically localizing lesions from CXR is apromising way to alleviate radiologists' daily reading burden. However, CXRdatasets often have numerous image-level annotations and scarce lesion-levelannotations, and more often, without annotations. Thus far, unifying differentsupervision granularities to develop thoracic disease detection algorithms hasnot been comprehensively addressed. In this paper, we present OXnet, the firstdeep omni-supervised thoracic disease detection network to our best knowledgethat uses as much available supervision as possible for CXR diagnosis. Besidesfully supervised learning, to enable learning from weakly-annotated data, weguide the information from a global classification branch to the lesionlocalization branch by a dual attention alignment module. To further enhanceglobal information learning, we impose intra-class compactness and inter-classseparability with a global prototype alignment module. For unsupervised datalearning, we extend the focal loss to be its soft form to distill knowledgefrom a teacher model. Extensive experiments show the proposed OXnet outperformscompetitive methods with significant margins. Further, we investigateomni-supervision under various annotation granularities and corroborate OXnetis a promising choice to mitigate the plight of annotation shortage for medicalimage diagnosis.