DeformableGAN: Generating Medical Images With Improved Integrity for Healthcare Cyber Physical Systems

• Published in: IEEE transactions on network science and engineering Vol. 10; no. 5; pp. 2584 - 2596
• Main Authors: Shen, Zhangyi, Ding, Feng, Jolfaei, Alireza, Yadav, Kusum, Vashisht, Sahil, Yu, Keping
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 01.09.2023; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: AI-driven cyber security; checkerboard artifacts; Convolution; Cyber-physical systems; Deep learning; Electronic mail; Generative adversarial networks; Gray-scale; Health care; healthcare; Image reconstruction; Image segmentation; Integrity; Machine learning; Medical diagnostic imaging; Medical electronics; Medical imaging; Medical research; Rehabilitation; Synthesis; Training
• ISSN: 2327-4697; 2334-329X
• DOI: 10.1109/TNSE.2022.3190765

The development of deep learning enables the production of new images via generative adversarial networks (GANs). The GANs have now been widely applied in the industry as well as academic research that brought tremendous progress to our community. Many researchers in medical imaging also introduced this novel technology for medical image reconstruction, segmentation, synthesis, etc. On the other hand, the GAN-generated images may also suffer from exhibiting unique textures, namely the checkerboard artifacts. For medical diagnosis and healthcare, such artifacts could bring negative impacts as they may distort information collected in medical images. Improper treatment and rehabilitation plans based on the disinformation of checkerboard artifacts could be harmful for patients and healthcare cyber physical systems. Thus, we investigate the checkerboard artifact synthesized during adversarial training in this paper. Based on the theoretical analysis, we propose a method for GANs to generate images without producing checkerboard artifacts. It could protect medical images preserving high integrity for healthcare cyber physical systems. Our experiments justify the efficiency of proposed method when associating with a variety of GANs for image synthesis. Also, we prove that it is feasible to detect GAN-generated images by tracing the checkerboard artifacts.