Automatic Pneumothorax Segmentation in Chest Radiographs Preprocessed by IEDSR Bone Suppression

High-contrast bone structure is the predominant interfering noise factor to diagnose various lung diseases including pneumothorax segmentation in chest radiographs. This paper applies a novel image preprocessing method of infrared-image enhanced deep super-resolution network (IEDSR) bone suppression to raise the accuracy and reliability of conventional automatic pneumothorax segmentation algorithms and projects. At the stage of image preprocessing, this paper adopts IEDSR model to detect and eliminate the bone structure and bone edge of posterior rib and clavicle on chest radiographs, and then enhances the integrity of soft tissue overlapped by the bone structure. Finally, the preprocessed image is fed to conventional automatic pneumothorax segmentation for region labeling of pneumothorax visualization in chest radiographs. Experimental results show, the novel image preprocessing method of IEDSR bone suppression is excellent at removing the high-contrast bone structure accurately and at reconstructing the subtle and low-contrast tissue structure of pneumothorax or lung marking. This novel bone suppression preprocessing method is verified to improve the accuracy and reliability of 3 conventional automatic pneumothorax segmentation algorithms and projects in chest radiographs on Dalin Tzu Chi (DTC) Hospital's dataset effectively and efficiently, better than various conventional bone suppression preprocessing methods.