Toward standardized evaluation of generative models for medical imaging

Generative models hold significant promise for advancing medical imaging by enabling privacy-preserving data sharing and augmenting limited datasets. However, their clinical translation remains constrained by insufficient validation, limited reproducibility, and the absence of standardized evaluation frameworks. This thesis contributes to addressing these challenges by introducing the Synthetic Images Metrics (SIM) Toolkit, an open-source framework for standardized assessment of 2D and 3D synthetic medical images across fidelity, diversity, and generalization. In addition, it presents PACGAN, a progressive auxiliary classifier GAN developed for conditional brain MR image synthesis, illustrating how architectural design can enhance data realism. Using brain MRI as a test case, two state-of-the-art generative models – StyleGAN2-ADA and Mediffusion – were systematically evaluated, revealing substantial performance variability due to random initialization, highlighting the necessity of multi-run benchmarking for fair and reproducible comparison. The influence of training set size on synthetic image quality was also examined. Collectively, these contributions establish methodological and practical foundations for the trustworthy and reproducible evaluation of generative models in medical imaging, promoting greater trust in synthetic data and supporting their future integration into research and clinical practice.