# Forecasting individualized progression of Alzheimer’s disease using structural MRI and population spatiotemporal priors

## TL;DR

This paper introduces a new method using MRI scans and population data to better predict how Alzheimer's disease progresses in individuals.

## Contribution

The novel approach combines progress maps with a GAN to improve individualized Alzheimer's progression forecasting.

## Key findings

- The pg-GAN model outperformed other models in predicting future MRI scans.
- Incorporating population-level progression data improved anatomical accuracy and prediction quality.
- The model showed better performance with reduced NRMSE and increased PSNR metrics.

## Abstract

Alzheimer’s disease (AD), the most common neurodegenerative disorder, involves the progressive loss of vulnerable neurons. Tracking its progression via structural magnetic resonance imaging (sMRI), which captures subtle brain anatomical changes, is vital for advancing diagnosis and treatment. Although generative models show promise in simulating disease progression by forecasting future magnetic resonance imaging (MRI) sequences, generating high-quality MRI with faithful anatomical structures remains challenging.

To narrow this gap, we proposed a progress map-guided generative adversarial network (pg-GAN) that leverages population-level longitudinal data to enhance individual-level prediction. First, progress maps were constructed by averaging intensity residuals between MRI scans acquired at different time points across a population, thereby preserving the comprehensive volumetric evolution of the brain over time. Then, the progress maps served as spatiotemporal priors and were embedded into a backbone generative adversarial network (GAN) via a proposed feature-wise fusion module (FFM) to predict future MRI for individuals.

We performed extensive experiments on 210 individuals with longitudinal MRIs from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset. The results demonstrated that our pg-GAN outperformed other conditioning models. The quantitative results showed that the normalized root mean squared error (NRMSE) decreased from 0.1623 to 0.1549, while the peak signal-to-noise ratio (PSNR) increased from 25.9353 dB to 26.3157 dB.

Incorporating group-level progression priors into the generative model can significantly improve the accuracy and anatomical fidelity of predicted MRIs, enhance the visualization of disease progression at the voxel level, and advance the development of precision treatment for AD.