Deep Generative Models in Brain MRI Synthesis for Alzheimer's Disease Research

Abstract

Alzheimer's Disease (AD) is a prevalent neurodegenerative disorder characterized by progressive cognitive decline. Early and precise diagnosis is crucial for effective prophylaxis and treatment. However, the scarcity of annotated medical imaging data, compounded by privacy restrictions and the high cost of acquiring detailed scans like brain Magnetic resonance imaging (MRI), which have proven to be essential in detecting early AD-related brain dysfunction changes, presents a significant challenge. To address the limited availability of such data, this thesis investigates the application of advanced deep generative models for synthesizing realistic 2D brain MRI slices specifically for AD research.
Here we introduce and explore the capabilities of state-of-the-art generative models and its variant, including Generative Adversarial Networks (GANs) and Diffusion Models (DMs), which have shown promise in creating synthetic medical images that closely resemble real patient data. By generating synthetic MRI images that are similar but not identical to the original inputs, these models show strong potential in enhancing the data pool available for AD research without compromising patient privacy. Through extensive experimentation and evaluation using mathematical metrics such as Kernel Inception Distance (KID) and Fréchet Inception Distance (FID), the study demonstrates that the DMs obtained the most effectiveness of these models in generating high-quality synthetic brain MRIs while Wasserstein GAN with gradient penalty presented great diversity. The achievements from three models are also displayed for an intuitive and subjective assessment and discussion. Overall, the results highlight the potential of these generative approaches to supplement existing datasets, enabling more robust and accurate models for AD diagnosis and research.
The thesis also discusses the importance of generating more complete 3D brain MRIs, which offer richer, voxel-to-voxel correlations, providing more informative insights for diagnosis, prediction, and disease staging. And the broader implications of generating integrating synthetic image data with clinical and biospecimen tabular data with biomarkers, suggesting a pathway toward creating comprehensive synthetic cohorts for AD research. While this work may not introduce groundbreaking innovations, it serves as a foundational step in the diagnosis, prediction, and staging of the disease.

Popular Abstract

Alzheimer's disease, one of the most common forms of dementia, is a progressive brain disorder that affects millions of people worldwide, leading to memory loss and cognitive decline. Early detection and effective research into its causes and treatments are critical, but studying the brain directly can be challenging. One of the most effective ways to study Alzheimer’s is through brain MRI scans, which allow researchers to examine the brain's structure and detect abnormalities. However, obtaining enough MRI data, especially from patients at various stages of Alzheimer’s, can be time-consuming, expensive, and sometimes impractical.
To address this, my research focuses on using deep generative models, a type of artificial intelligence (AI), to create synthetic brain MRI scans. These AI models are trained on existing MRI data and learn to create new, realistic images that closely resemble those of actual patients. By creating high-quality, realistic brain images, these AI models can help scientists test new methods for diagnosing Alzheimer’s, track disease progression, and explore potential treatments — all without relying limited real patient scans. Not only various advanced generative AI models were attempted in this thesis work, but also performance evaluation and discussion, which show that these models have the potential to artificially manufacture outputs that could easily be mistaken for real data. The goal is to provide researchers with more medical data, making it easier to study the disease. In the future, this could lead to earlier diagnosis, improved patient outcomes, and a deeper understanding of how the disease affects the brain. However, this work represents just a groundwork attempt for a much larger, more comprehensive project, since the current research primarily builds on existing techniques, and with some slight innovations, future advancements could be extensive.
In summary, this project serves as a stepping stone, highlighting the potential of AI in healthcare and laying the foundation for future research. In this thesis, we are focusing on applying current techniques. It opens up exciting possibilities for improving Alzheimer’s diagnosis, tracking disease progression, and even contributing to the development of new treatments. In the long run, these advancements could significantly accelerate the fight against Alzheimer’s disease and improve patient care.

• author: Zhao, Ruoyi
• supervisor: Ida Arvidsson; Filip Winzell; Linda Karlsson; Karl Åström
• organization: Mathematics (Faculty of Engineering)
• course: FMAM02 20241
• year: 2024
• type: H2 - Master's Degree (Two Years)
• subject: Mathematics and Statistics
• keywords: generative models, Alzheimer's Disease, Magnetic Resonance Imaging(MRI), Alzheimer's Disease Neuroimaging Initiative(ADNI), Generative Adversarial Networks(GANs), Diffusion Models(DMs), Wasserstein GAN with Gradient Penalty(WGAN-GP), Denoising Diffusion Probabilistic Models(DDPM).
• publication/series: Master’s Theses in Mathematical Sciences
• report number: LUTFMA-3559-2024
• ISSN: 1404-6342
• other publication id: 2024:E68
• language: English
• id: 9174608
• date added to LUP: 2024-09-18 15:27:44
• date last changed: 2024-09-18 15:27:44