From Structure to Function and Back Again: A GAN-Guided Diffusion Framework for Generating Clinically Meaningful...

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Biomedical Data Science

Background:

• Multimodal brain imaging offers rich insights but is often hampered by missing data.
• Conventional methods for handling missing data can introduce bias or discard valuable information.
• Generative models show promise for synthesizing missing neuroimaging modalities.

Purpose of the Study:

• To develop a novel generative framework for cross-modality translation in brain imaging.
• To synthesize missing T1-weighted magnetic resonance imaging (MRI) and functional network connectivity (FNC) data.
• To enable robust multimodal brain imaging analysis even with incomplete datasets.

Main Methods:

• Introduced a Generative Adversarial Network (GAN)-guided diffusion framework.
• Integrated conditional diffusion modeling, adversarial learning, and cycle-consistency.
• Enabled training with both paired and unpaired multimodal brain imaging data.

Main Results:

• Achieved superior performance in synthesizing T1-weighted MRI, with high PSNR (24.95) and SSIM (0.86).
• Demonstrated improved correlation (0.65) with real functional network connectivity data.
• Showcased the model's ability to capture clinical group variability without diagnostic labels.

Conclusions:

• The GAN-guided diffusion framework effectively synthesizes realistic and clinically meaningful T1-weighted MRI and FNC data.
• This approach enhances multimodal brain imaging analysis, particularly for conditions like Alzheimer's disease.
• The method facilitates downstream analysis and biomarker discovery by generating high-quality synthetic modalities.