Comparative flow-based brain MRI generation trained on the ADNI dataset conditioned on diagnostic labels

Alzheimer’s Disease (AD) is the most common cause of dementia, and structural T1-weighted MRI represents one of its most established imaging biomarkers. The development of deep learning models for medical research is often limited by the scarcity, class imbalance and privacy constraints of medical imaging datasets; synthetic data generation has emerged as a promising strategy to address these limitations. This thesis compares two generative approaches for the synthesis of 3D T1-weighted brain MRI volumes: the Denoising Diffusion Probabilistic Model (DDPM) and Conditional Flow Matching with Optimal Transport (CFM + OT). Both models were trained on T1-weighted volumes from the ADNI dataset, preprocessed through a dedicated pipeline, and conditioned on both anatomical head masks and diagnostic labels (Cognitively Normal, Mild Cognitive Impairment, Alzheimer’s Disease). To ensure a fair comparison, both models share the same underlying 3D U-Net architecture, input data, and reproducible components of the sampling procedure, while hyperparameters were optimized independently for each model. The synthetic image quality was compared between models and evaluated against a real-data baseline using four complementary metrics: Peak Signal-to-Noise Ratio (PSNR), Multi-Scale Structural Similarity Index Metric (MS-SSIM), Maximum Mean Discrepancy (MMD) and Fréchet Inception Distance (FID), combined with paired and non-parametric statistical tests. Across all four metrics and all diagnostic groups, the CFM + OT model consistently outperformed the DDPM; for PSNR and MS-SSIM, this difference was also statistically significant. The CFM + OT model also offered a substantial practical advantage, reducing training time by approximately 40% and sampling time by approximately 26%, compared to the DDPM. Despite this improvement, a non-negligible gap between both models and the real-data baseline remains across all metrics.