55 Facilitating Valid Statistical Inference in Biomedical Image Synthesis

Oregon Convention Center 

Image synthesis is an important and growing field of research fueled by rapid progress in image-based artificial intelligence and has been employed in neuroimaging, multiplexed immunofluorescence (MxIF), and imaging spatial transcriptomics. The number of publications on image synthesis has nearly tripled in the last decade, but there is no evaluation of the consequence of using it in medical research. Currently, biomedical image synthesis 1) does not include relevant clinical information, and 2) fails to provide statistical uncertainty. As postulated by multiple imputation theory, neglecting either issue can lead to invalid downstream statistical analysis. In this paper, we systematically examine these issues in state-of-the-art image synthesis algorithms with real-world imaging data. We demonstrate that 1) current imputation tools often lead to biased point estimates and anti-conservative standard errors, and 2) such issues can be alleviated by simple, post-hoc augmentation steps derived from multiple imputation literature. This work is pioneer in highlighting invalid findings on synthesized biomedical imaging data, and providing expeditious solutions.

Statistical imaging

Machine learning

Spatial genomics

Multiple imputation 

Section on Statistics in Genomics and Genetics