Understanding and mitigating data replication in medical synthetic image generation using diffusion models

Recorded talk

Recent advancements in diffusion models have significantly improved the quality of generated images. This progress is particularly promising in the medical domain, where data scarcity, especially for rare diseases and cases, is a common challenge. While diffusion models generate diverse images, they also exhibit replication of images from the training dataset source. This issue could be crucial in the medical domain where safeguarding patient privacy is non-negotiable. Our study delves into the issue of image replication by diffusion models in medical imaging, examining its prevalence and implications for patient privacy. We first extensively study the types of replication that could occur in synthetic data and introduce methods by which one can identify and quantify the extent of replication present. We then introduce mitigation protocols and strategies that can be implemented before distributing synthetic data to ensure patient confidentiality and proactively prevent instances of replication. We further examine how replicated and non-replicated synthetic data after our proposed mitigation strategy influence medical image classification tasks across X-ray, Ultrasound, and CT modalities.

Vishal M. Patel is an associate professor of electrical and computer engineering and a member of the Vision and Image Understanding Lab. His research interests are focused on computer vision, machine learning, image processing, medical image analysis, and biometrics. He has received a number of awards including the 2021 IEEE Signal Processing Society (SPS) Pierre-Simon Laplace Early Career Technical Achievement Award, the 2021 NSF CAREER Award, the 2021 IAPR Young Biometrics Investigator Award (YBIA), the 2016 ONR Young Investigator Award, and the 2016 Jimmy Lin Award for Invention. Patel is an associate editor of the IEEE Transactions on Pattern Analysis and Machine Intelligence journal and chairs the conference subcommittee of IAPR Technical Committee on Biometrics (TC4).