Synthetic Brain Dataset: High-Fidelity Medical Image Generation 🧠

Apr, 2025

The goal is to generate highly realistic brain MRI scans with fine-grained pathological details such as tumor presence, location, size, shape, intensity, and brain orientation.

Unlike general text-to-image generation, we work with extremely precise medical descriptions — for example: "1: tumor: yes; location: pituitary; size: large; shape: regular; intensity: bright; orientation: sagittal; general description: Brain MRI in sagittal view showing large pituitary tumor with abnormal enhancement."

We use vision-language models to produce detailed annotations for existing medical images, and fine-tune diffusion models to synthesize new, diverse examples guided by these structured prompts. This enables us to simulate rare conditions and specific anatomical variations, with potential applications in data augmentation, privacy-preserving model training, and medical education.

Part of the Advanced Deep Learning in Computer Vision course project supervised by Changlu Guo at DTU.

Code and more details coming soon!