Score-based generative diffusion models to synthesize full-dose FDG brain PET from MRI in epilepsy patients

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Physics

Background:

• Fluorodeoxyglucose (FDG) PET/MRI is crucial for epilepsy evaluation, assessing brain structure and metabolism.
• Current methods involve significant radiation dose, particularly concerning for young epilepsy patients.
• Advanced generative AI, specifically diffusion models, offers potential for synthesizing PET images from MRI data, reducing radiation exposure.

Purpose of the Study:

• To compare deep learning models, including diffusion models, for MRI-to-PET image translation in epilepsy.
• To evaluate the performance of models using different MRI contrasts (T1w, T2 FLAIR) and ultralow-dose PET data.
• To assess the clinical relevance of synthesized PET images for evaluating metabolic asymmetry in epilepsy.

Main Methods:

• Simultaneous PET/MRI data from 52 epilepsy subjects were used for training, validation, and testing.
• Three deep learning models were compared: Score-based Generative Diffusion Models (SGM-KD, SGM-VP) and a Transformer-U-net.
• Models were evaluated using standard image processing metrics and clinical congruency measures for hemispheric asymmetry.

Main Results:

• The SGM-Karras Diffusion model achieved the best results synthesizing PET from T1w and T2 FLAIR MRI alone.
• Incorporating 1% ultralow-dose PET data significantly improved all models, making them quantitatively and visually interchangeable.
• All tested models demonstrated high accuracy in synthesizing full-dose FDG-PET using MRI and ultralow-dose PET.

Conclusions:

• Deep learning diffusion models show promise for pure MRI-to-PET synthesis in epilepsy imaging.
• Combining MRI with ultralow-dose PET enables accurate PET image synthesis across tested deep learning models.
• This AI-driven approach has the potential to substantially reduce or eliminate radiation dose for epilepsy patients undergoing PET/MRI evaluation.