Diffusion model enables quantitative CBF analysis of Alzheimer's Disease

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Cerebral blood flow (CBF) measured by arterial spin labeling (ASL) is a key biomarker for Alzheimer's Disease (AD).
• The Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset contains ASL data from multiple vendors, but missing M0 images from Siemens scanners hinder CBF quantification.
• Imputing missing M0 images is crucial for leveraging large-scale ASL datasets in AD research.

Purpose of the Study:

• To develop and validate a generative diffusion model for imputing missing M0 images in Siemens ASL data.
• To enable accurate CBF quantification for AD research using the ADNI dataset.
• To compare the performance of imputed CBF data with acquired CBF data in differentiating AD stages and predicting AD.

Main Methods:

• A conditional latent diffusion model was trained to generate M0 images.
• The model was validated using image similarity metrics, CBF quantification accuracy, and physical model consistency on an in-house dataset (N=55).
• The validated model was applied to the ADNI Siemens dataset (N=211) to impute M0 images for CBF calculation and compared with GE acquired data.

Main Results:

• The diffusion model generated M0 images with high fidelity (SSIM=0.924±0.019, PSNR=33.348±1.831).
• Imputed CBF data showed minimal bias (mean difference=1.07±2.12ml/100g/min) and comparable differentiation patterns across AD stages and classification performance to acquired data.
• Generated CBF data improved AD stage classification accuracy compared to qualitative perfusion data.

Conclusions:

• Generative diffusion models are effective for imputing missing M0 modalities in ASL data.
• This approach enables robust CBF quantification in large-scale neuroimaging studies, particularly for Alzheimer's Disease.
• The method holds significant potential for advancing AD research by unlocking the full value of existing ASL datasets.