Unsupervised MR harmonization by learning disentangled representations using information bottleneck theory

• DOI: 10.1016/j.neuroimage.2021.118569
• Publication Date: 2021-09-08T08:12:40Z
• Authors: Lianrui Zuo; Blake E. Dewey; Yihao Liu; Yufan He; Scott D. Newsome; Ellen M. Mowry; Susan M. Resnick; Jerry L. Prince; Aaron Carass
• Keywords: Harmonization; Sample (material)

In magnetic resonance (MR) imaging, a lack of standardization in acquisition often causes pulse sequence-based contrast variations MR images from site to site, which impedes consistent measurements automatic analyses. this paper, we propose an unsupervised image harmonization approach, CALAMITI (Contrast Anatomy Learning and Analysis for Intensity Translation Integration), aims alleviate multi-site imaging. Designed using information bottleneck theory, learns globally disentangled latent space containing both anatomical information, permits harmonization. supervised methods, our approach does not need sample population be imaged across sites. Unlike traditional approaches suffer geometry shifts, better preserves anatomy by design. The proposed method is also able adapt new testing with straightforward fine-tuning process. Experiments on acquired ten sites show that achieves superior performance compared other approaches.