A Study on Training Data Augmentation Technique for Segmentation AI Using CycleGAN

Abstract

Japan has one of the lowest autopsy rates among developed countries, which is thought to be mainly due to a shortage of forensic scientists. Therefore, if the cause of death can be determined by using artificial intelligence on postmortem computed tomography (CT) images, this could be a solution to the shortage of forensic scientists and image readers.In this study, we have been working on basic research to realize a system that enables prediction of organ weights, which is important information for identifying the cause of death, from postmortem CT images. However, postmortem changes are large in the lung region, making it difficult to secure sufficient data to improve the accuracy of the learning model.Therefore, in this study, we attempted to extend the data by pseudo-generating postmortem CT images from biological CT images using CycleGAN, and we report on these attempts.