Generative AI and Counterfactual Image Generation for Interpretable Alzheimer’s Disease Detection - ABSTRACT Although neuroimaging provides unprecedented resources for understanding brain disease, it is still extremely challenging to distinguish brain changes due to Alzheimer’s disease (AD) from the complex pattern of brain atrophy that accompanies normal aging, making automatic detection of AD very difficult. A promising solution is to use advances in generative AI - in particular, denoising diffusion probabilistic models (DDPM) applied to images - to encode the high-dimensional variations in the living brain, learn the normal range and statistical distribution of brain variations - conditional on a person’s age, sex, and other risk factors. Standard statistical mapping detects AD based on brain MRI or amyloid PET (using SPM and TBM) relying on mass-univariate statistics that encode the variations in one pixel or location in the brain, without considering holistic information on the patterns of deviations from the normal range of variation. Addressing this, we create a powerful generative AI method to assist in the diagnostic evaluation of AD, to (1) highlight pathology; (2) provide personalized probabilistic disease models (digital twins) to predict future trajectories; (3) create controllable simulations to study factors affecting progression. We propose an innovative approach based on conditional latent diffusion models (LDM) and denoising diffusion probabilistic models (DDPMs) to provide insight into AD effects on individual brain anatomy. First, our generative models will learn from well-known neuroimaging datasets paired with diverse subject-specific data (age, sex, scanning protocol, clinical diagnosis of AD, and amyloid and tau positivity defined using PET, CSF or blood markers). We will train diffusion models on real 3D T1-weighted brain MRI scans and create conditional generative models to encode how the data distributions depend on clinical diagnosis of AD, and amyloid and tau positivity. We will evaluate the ability of the diffusion models to conditionally sample MRIs, using a 3D CNN-based disease classifier trained on real MRIs. Next, we will use implicit classifier-free guidance to alter the conditioning of a person’s scan into its disease-free counterfactual image, while preserving subject-specific image details. A corresponding personalized disease map will be generated to identify possible AD effects in the brain. We will evaluate methods to condition synthetic data on age, sex, and scanning protocol, generating realistic, diverse data to support interpretable AI- based clinical decision-making and new maps of neurodegenerative disease effects for genetic and risk factor studies. Our revised R21 responds to excellent reviewer advice and includes new preliminary data for model interpretability, hallucination mitigation techniques, a radiologist collaborator for error flagging, enhanced dataset diversity, additional training data, and comparative benchmarking. We will compare our DDPM maps of AD effects with standard SPM-based methods (voxel- and tensor-based morphometry) to assess sensitivity to clinically-defined AD and amyloid/tau positivity.
