Are Medical Image Generative Models Biologically Trustworthy?

Suhyun Ahn, Wonjung Park, Jinah Park; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025, pp. 644-654

We investigate the biological trustworthiness of deep generative models in capturing intrinsic variability in medical imaging, with a focus on neurodegenerative disease and sex dimorphism. While recent generative models for medical images have demonstrated high perceptual quality, their ability to preserve biologically meaningful variations remains underexplored. In this work, we systematically evaluate two representative generative models conditioned on sex and Alzheimer's disease cognitive states. Using real brain MRI data, we first characterize sex- and disease-related anatomical variability, then assess whether the generated synthetic images accurately reflect these patterns. Despite producing visually plausible outputs, both models exhibit substantial limitations in capturing structural differences across key brain regions, indicating low biological fidelity. Our findings highlight critical gaps in current generative approaches and underscore the need for biologically grounded evaluation metrics to ensure clinically meaningful and equitable synthetic data generation.