Detection Of Visual Signals For Pneumonia In Chest Radiographs Using Weak Supervision

Pneumonia usually manifests as areas of increased opacity in chest radiographs. Diagnosing pneumonia often requires a review of a chest radiograph by highly trained specialists. Automatic identification of regions of interest in chest radiographs could be an initial step in prioritizing radiology worklist for radiologists to review. In this paper, we show how to exploit class-activation mapping and weak supervision to automatically localize regions in chest radio-graphs exhibiting signs of pneumonia. For this research project, we evaluate a dataset of 30,000 chest radiographs collected by the Radiology Society of Northern America (RSNA). Data was annotated by six board certified radiologists who provided bounding boxes to specify lesion location. We investigated the utility of weak supervision for detecting visual signs of pneumonia in chest radiograph (CXR). We formulate the problem as one of inference in binary pneumonia classification. Binary classification only requires a binary label indicating the presence of pneumonia as opposed to bounding boxes or detailed annotations outlining the regions of interest. We compared the regions of interest identified by our approach and a data-set of chest x-rays with detained ROI bounding box annotations prepared by the Radiological Society of North America (RSNA) in collaboration with the US National Institutes of Health, We show that our approach achieves good correlation intersection-over-union with the radiologists ground truth annotations from the (RSNA) dataset. Weakly supervised learning can improve the data curation burden by using weak labels in training for identification of pneumonia. These findings have broader implications, as annotating and labeling large datasets for medical imaging applications is a difficult task, and is often the bottleneck in efforts to apply advances in machine learning and deep learning.