Learning the Latent Heat Diffusion Process through Structural Brain Network from Longitudinal $\beta$-Amyloid Data

Md Asadullah Turja, Guorong Wu, Defu Yang, Martin Andreas Styner

Keywords: Heat Diffusion, Florbetapir PET, Structural Brain Network, Alzheimer's Disease, Graph Neural Network, Graph Embedding, Graph u-net, Amyloid Progression Pattern, $\beta$-Amyloid Pathology.

TL;DR: A graph u-net based architecture to learn hidden diffusion mechanism in Alzheimer's brain from longitudinal amyloid data.

Abstract: The excessive deposition of misfolded proteins such as amyloid-$\beta$~(A$\beta$) protein is an aging event underlying several neurodegenerative diseases. Mounting evidence shows that the spreading of neuropathological burden has a strong association to the white matter tracts in the brain which can be measured using diffusion-weighted imaging and tractography technologies. Most of the previous studies analyze the dynamic progression of amyloid using cross-sectional data which is not robust to the heterogeneous A$\beta$ dynamics across the population. In this regard, we propose a graph neural network-based learning framework to capture the disease-related dynamics by tracking the spreading of amyloid across brain networks from the subject-specific longitudinal PET images. To learn from limited (2 – 3 timestamps) and noisy longitudinal data, we restrict the space of amyloid propagation patterns to a latent heat diffusion model which is constrained by the anatomical connectivity of the brain. Our experiments show that restricting the dynamics to be a heat diffusion mechanism helps to train a robust deep neural network for predicting future time points and classifying Alzheimer's disease brain.

Registration: I acknowledge that publication of this at MIDL and in the proceedings requires at least one of the authors to register and present the work during the conference.

Source Code Url: https://github.com/mturja-vf-ic-bd/Diffusion_u_net/tree/master

Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.

Paper Type: both

Primary Subject Area: Learning with Noisy Labels and Limited Data

Secondary Subject Area: Unsupervised Learning and Representation Learning