Chest X-ray (CXR) is one of the most commonly prescribed medical imagingprocedures, often with over 2-10x more scans than other imaging modalities suchas MRI, CT scan, and PET scans. These voluminous CXR scans place significantworkloads on radiologists and medical practitioners. Organ segmentation is acrucial step to obtain effective computer-aided detection on CXR. In this work,we propose Structure Correcting Adversarial Network (SCAN) to segment lungfields and the heart in CXR images. SCAN incorporates a critic network toimpose on the convolutional segmentation network the structural regularitiesemerging from human physiology. During training, the critic network learns todiscriminate between the ground truth organ annotations from the maskssynthesized by the segmentation network. Through this adversarial process thecritic network learns the higher order structures and guides the segmentationmodel to achieve realistic segmentation outcomes. Extensive experiments showthat our method produces highly accurate and natural segmentation. Using onlyvery limited training data available, our model reaches human-level performancewithout relying on any existing trained model or dataset. Our method alsogeneralizes well to CXR images from a different patient population and diseaseprofiles, surpassing the current state-of-the-art.
展开▼