Keywords: CT Synthesis, CycleGAN, Simulation- Based Deep Learning
Abstract: Deep learning in medical imaging is often limited by the availability of training data with sufficient quality. The synthesis of image data offers a solution to this data shortage. Here, we use the CycleGAN network architecture to synthesize axial CT slices based on anthropomorphic body phantoms. We investigate the influence of an identity loss and a gradient difference loss function on the image quality of the synthesized data. We evaluate the synthesized images with respect to anatomical accuracy and realistic CT noise properties. The additional loss functions improved the preservation of edges and anatomical structures compared to the original CycleGAN loss, without deteriorating the noise quality of the synthetic image.
Code Of Conduct: I have read and accept the code of conduct.
Remove If Rejected: Remove submission from public view if paper is rejected.