Facilitating Valid Statistical Inference in Biomedical Image Synthesis

Abstract

Generative AI for medical image synthesis is increasingly used to impute missing data and augment biomedical datasets. Yet despite achieving high computational similarity to their real-world counterparts, synthetic images remain uncommon in clinical research practice, in part because of concerns about their biological validity. Here, we evaluate whether synthetic images support reliable downstream statistical analysis and biological inference in magnetic resonance imaging (MRI) and multiplexed immunofluorescence (mIF) imaging. We find that associations between patient phenotypes and synthesized images can be severely attenuated. This is because synthesis methods typically omit clinically relevant covariates and therefore generate images that, although visually realistic, mask key biological differences. Furthermore, naive analyses treat synthesized images as if they were observed, ignoring the statistical uncertainty introduced by synthesis. Drawing on multiple imputation theory, we introduce Phenotype Aware Multiple Synthesis (PAMS), an analysis framework that requires no model retraining and instead adjusts synthesized output to recover valid biological estimates with appropriately calibrated uncertainty. Applied to MRI and mIF data, PAMS reduced estimation bias and restored nominal confidence interval coverage. Because it is compatible with existing synthesis pipelines, PAMS offers a practical and rigorous route to statistically and biologically valid use of synthesized images in biomedical imaging research.