Region-dependent Graph Representation Learning on Brain Morphable Meshes
Link to paper: https://openreview.net/pdf?id=J4JWTCq14u
- Abstract:
We propose a geometric deep learning framework, using a novel mesh-based pooling module called RegionPool, to classify MCI subtypes (EMCI/LMCI) and investigate regional brain morphological changes. Our method shows superior classification capability and reveals that current MCI subtypes fail to identify diverse patterns of cortical atrophy during the MCI stage. The generated class activation maps (CAMs) provide visual evidence supporting our model’s decisions and align with atrophy patterns reported in relevant literature. This approach is important for preventive interventions in slowing down the progression from MCI to Alzheimer’s disease (AD).
- Methods:
Construction of graphs: A single mesh instance will contain 47,616 vertices, including 32,768 (69%) from the cortical structure and 14,848 (31%) from the subcortical structure.
Model structure: Het-BMAT is built upon the GAT convolution, our proposed RegionPool and a combination of hetero and normal linear layer.
Region-dependent Pooling on Brain Mesh Surface: The process of mesh simplification and building the mesh hierachies.
- Experiment & Results
4,072 T1-weighted MRI volumes from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset are used for conducting experiments.
During the experiment, we made the following observations: Firstly, we found that morphological changes occurring in the subcortical structures play a more critical role in the classification process. Secondly, it was noted that empirically derived MCI may not adequately capture the heterogeneity of cortical atrophy at the MCI stage. And finally, there was a significant involvement of the amygdala, left putamen, and left thalamus in the classification, indicating a remarkable asymmetry in the class activation maps (with left lateralization).
- Conclusion:
Het-BMAT together with RegionPool demonstrates a superior MCI subtype classification capability using only geometric information.
Revealing the limitation of using a single neuropsychological score to capture the heterogeneous patterns of cortical atrophy at the MCI stage.
The CAMs) indicate the left lateralization of brain atrophy during the process of AD and are consistent with the morphological changes reported in the related literature.