Flow-Matching for Generative Medical Imaging

This project was realized during a research internship at TUM Bioinformatics Professorship (BIT) as part of my M.Sc. Chemical Biotechnology degree.

The goal was to implement, train and evaluate Flow-Matching and Diffusion (Probabilistic) Models using U-Net and VisionTransformer architectures. Both were used for un- and conditional synthesis of MRI images via Classifier Free Guidance.

The models were trained on the Brain Tumor MRI Dataset and Smithsionian Butterflies for additional validation.

Results

• Flow-Matching Models outperform DDPMs in terms of sample quality and inference time. The former generates images containing fewer aberrations and artifacts.
• Regardless of model class, U-Net outperforms ViT and enables the network to generate images of higher quality.
• The ViT was incapable of training on the preprocessed Smithsonian Butterflies Dataset as it contained a small number of training samples.
• Increasing the guidance scale to extremely high values yields highly saturated and distorted samples (as can seen below):

References & Acknowledgements

The Flow-Matching code was heavily inspired by MIT's 6.S184 (2025) course.

• MRI Dataset
• Smithsonian Butterflies Dataset
• Denoising Diffusion Probabilistic Models paper and code
• Flow-Matching Guide and Code paper and code