Synthetic Cardiac MRI Image Generation using Deep Generative Models

In this project, students will explore and evaluate deep learning models for generating synthetic CMRI images using limited real samples. They will implement and compare generative approaches such as Variational Autoencoders (VAEs), Generative Adversarial Networks (GANs), and Diffusion Models. The project may also explore few-shot learning strategies and inpainting techniques to enhance synthetic image quality and diversity.

This project is derived from the EU-funded SEARCH project (https://ihi-search.eu), which aims to develop trustworthy synthetic medical data to support AI research in healthcare. One major aspect of SEARCH is the generation of high-quality synthetic radiological images, including Cardiac Magnetic Resonance Imaging (CMRI).

In this project, students will explore and evaluate deep learning models for generating synthetic CMRI images using limited real samples. They will implement and compare generative approaches such as Variational Autoencoders (VAEs), Generative Adversarial Networks (GANs), and Diffusion Models. The project may also explore few-shot learning strategies and inpainting techniques to enhance synthetic image quality and diversity.

Goals / learning outcomes

Students will:

• Use public CMRI datasets or simulate data for experimentation.
• Implement and compare 1–2 generative models for CMRI image synthesis.
• Evaluate generated images using perceptual similarity metrics (e.g., SSIM, PSNR, FID).
• Optionally explore inpainting methods to add or remove abnormalities.
• Reflect on the utility, privacy, and diversity aspects of the synthetic data.

This project contributes to SEARCH’s broader goal of developing realistic and privacy-preserving synthetic radiological data for safe and effective AI applications in cardiology.

Medical image data is critical for developing AI tools, yet often inaccessible due to privacy concerns. This project enables students to apply advanced generative models to address this limitation, with direct relevance to cardiac diagnostics and radiology.

Requirements

• Basic understanding of computer vision and deep learning
• Familiarity with convolutional neural networks and Python programming
• Experience with image processing or evaluation metrics is a plus

Students will use publicly available datasets. Ethical handling of data and reproducible code practices are required, aligning with the open science goals of the SEARCH project.

Supervisors

• Vajira Thambawita
• Molle Maleckar
• Pål Halvorsen

Collaboration partners

This project is part of SEARCH (Synthetic hEalthcare dAta goveRnanCe Hub), a multi-disciplinary initiative focused on creating synthetic healthcare data and facilitating secure data sharing across the biomedical ecosystem. Read more about SEARCH here.