Multimodal Generative Modeling for DaT Scan Reconstruction in Parkinson's Disease

Area of Science:

• Nuclear Medicine
• Neuroimaging
• Artificial Intelligence

Background:

• Generating synthetic medical data that mirrors real-world distributions while protecting privacy is challenging.
• Nuclear medicine research faces data sharing hurdles due to regulations and ethics.
• Accurate DaT scans are crucial for Parkinson's disease diagnosis and research.

Purpose of the Study:

• To develop a multimodal deep learning model for reconstructing ¹²³I-FPCIT SPECT (DaT) scans.
• To enable the future synthesis of privacy-preserving, statistically representative medical data.
• To validate the model's performance on diverse patient cohorts.

Main Methods:

• A multimodal deep learning framework was designed to leverage co-registered T1-weighted MRI and DaT scans.
• Extensive experiments were conducted on the Parkinson's Progression Markers Initiative (PPMI) dataset.
• The dataset included healthy controls, Parkinson's disease (PD) patients, and subjects without dopaminergic deficits (SWEDD).

Main Results:

• The proposed framework successfully reconstructed DaT images that closely preserved clinical signal distributions.
• Minimal intensity discrepancies and unbiased contrast-to-noise ratios were observed.
• Robust region-based analyses confirmed the model's effectiveness across different subgroups.

Conclusions:

• The deep learning model demonstrates feasibility for reconstructing DaT scans from multi-contrast inputs.
• This approach can enhance the generation of synthetic data for neurodegenerative research.
• The framework supports large-scale data augmentation, aiding early PD detection and disease progression studies.