Reduction of Unsharpness Caused by Patient Motion on Radiography using Deep Learning

In radiographic examination, in case of the patient’s body movement or incomplete breath-holding, radiography was retaken again. Therefore, we investigated in U-net, Cycle-GAN, and UNIT, and developed a method for remove blur on radiographs with higher accuracy. The database used in this study consists of 39 cases of pair of radiography (blurred image and reference image) at the Eastern Chiba Medical Center, in which the radiography was retaken due to body movement or incomplete breath-holding. Next, for data augmentation, the number of training images was increased by rotation and inversion, the blurred image was input to U-net, Cycle-GAN and UNIT to generate no-blurred image. Comparing with the reference image, four radiological technologists performed visual evaluation of the test images on 5-point scale. Blurred score, retention of anatomy, contrast change on visual evaluation in UNIT was higher than this in U-net and Cycle-GAN, and there was a statistically significant difference (P