Graph Neural Network Reveals the Cortical Morphology of Local Brain Aging in Normal Cognition and Alzheimer's Disease

Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.

Ines: We're looking at the paper: "Graph Neural Network Reveals the Cortical Morphology of Local Brain Aging in Normal Cognition and Alzheimer's Disease".

Marcus: Here's the gist of the paper. Graph Neural Network Reveals the Cortical Morphology of Local Brain Aging in Normal Cognition and Alzheimer’s Disease is a study introducing a graph neural network to estimate local brain aging from cortical morphology, which is significant because it provides a biologically interpretable framework for quantifying local brain aging patterns across normal cognition and Alzheimer’s disease.

Ines: Estimating brain age (BA) from T1-weighted magnetic resonance images (MRIs) provides a powerful framework for quantifying anatomical brain aging Whereas global BA (GBA) summarizes overall brain health, local BA (LBA) provides cortically specific patterns of aging at the subject level Model Introduction and Methodology The paper introduces a graph neural network (GNN) that uses morphometric features—cortical thickness, surface area, curvature, gray/white matter intensity ratio (GWR), sulcal depth—to estimate LBA across the cortical surface at high spatial resolution Trained on cortical surface meshes extracted from the MRIs of cognitively normal (CN) adults (N = 14,423), our model achieves lower mean absolute error (MAE) than the existing state-of-the-art while identifying more biologically plausible patterns of aging in Alzheimer’s disease (AD) on the ADNI dataset The approach conceptualizes the cortical surface, derived from T1w MRI, as a three-dimensional surface mesh, with morphometric features mapped to each vertex along this mesh This mesh and its features are used as input to the model, facilitating the use of explainability techniques—such as feature ablation—to identify the morphological underpinnings of local aging Data Preparation and Preprocessing The data utilized T1w MRI scans aggregated across multiple sources to enhance the generalizability of our model and findings The training sample was comprised of CN adults from the UK Biobank (UKBB), National Alzheimer’s Coordinating Center (NACC), and Information eXtraction from Images (IXI) datasets, totaling 14,423 scans spanning a broad CA range Preprocessing involved using FreeSurfer (FS) to extract high-resolution cortical surface features, including cortical thickness, sulcal depth, curvature, surface area, and GWR The FS pipeline included intensity normalization, skull stripping, bias field correction, and segmentation of brain tissue into white matter (WM), gray matter (GM), and cerebrospinal fluid (CSF) To facilitate sample-level learning, each mesh was projected onto a standardized, hemisphere-specific cortical atlas Model Architecture and Training The model follows a hierarchical U-Net architecture designed for hemisphere-concatenated cortical meshes in atlas space It consists of five blocks arranged across three FS atlas resolutions: ico6 (V6 = 81,924 vertices), ico5 (V5 = 20,484 vertices), and ico4 (V4 = 5,124 vertices) Within each block, feature representations are derived or refined using a single graph convolutional network layer (GCNConv) The input block receives five morphometric features per vertex (F0 = 5), while the number of derived feature channels in subsequent blocks (F1, F2) is treated as a tunable hyperparameter Model hyperparameters were optimized using an 80/20 subject-level training/validation split derived from the training data The model was optimized using the AdamW optimizer with feature sizes of F1 = 256 and F2 = 512, a batch size of 8, a learning rate of 0.0005, and a weight decay of 0.0003 for 60 epochs Key Findings on Aging Patterns The model achieved an average MAE of 5.63 y for the CN cohort prior to bias correction when evaluated on the external, independent, ADNI dataset In CNs, the largest positive effects were observed in visual, parietal, and prefrontal association cortices Individuals with MCI exhibited widespread positive ΔLBAGs across regions and networks In AD subjects, the largest regional ΔLBAGs were observed in the inferior temporal gyrus (2.36 y) and the parahippocampal gyrus (2.04 y) Feature ablation showed that surface area demonstrated its strongest negative effects throughout the visual and parietal association cortices Curvature and GWR exhibited the strongest effects in the inferior temporal gyrus, with a strong effect in the parahippocampal gyrus Clinical Correlations and Interpretation Regressing LBAGs on cognitive scores showed no significant associations between LBAGs and cognitive scores for the CN cohort By contrast, several cognitive tests were associated with regional LBAGs for the AD cohort, including the Montreal Cognitive Assessment (MoCA; β = 1.41) and Alzheimer’s Disease Assessment Scale-11 (ADAS11; β = 1.25) The inferior temporal gyrus exhibited larger effect sizes than those observed in the precentral gyrus, and demonstrated additional significant associations with RAVLT Percent Forgetting and DSST scores in the AD cohort Comparison to Existing Models The GNN model achieved a 0.31-year lower MAE than the LBA-SotA (MAE = 5.94 y)

37: The GNN model indicated substantially greater aging in the temporal lobes of AD subjects, whereas the LBA-SotA identified approximately uniform LBAG differences across all lobes This finding suggests that our GNN model may be better suited for LBA estimation than CNN alternatives The model identified association cortices as particularly vulnerable in normative aging, with the parahippocampal gyrus implicated in both MCI and AD, highlighting its potential as an early marker of disease progression Feature-level analyses further suggest that different morphometric properties capture distinct aspects of brain aging, with surface area more sensitive to normal aging than to AD-related decline The persistent involvement of the parahippocampal gyrus across MCI and AD supports its role as a sensitive marker of disease progression Future studies may benefit from the implementation of models that combine accurate local representations with more explicit multiscale organization Limitations A limitation of our study lies in the composition of the training and testing sets, as the training set was skewed toward older adults and not fully age-matched A more theoretical limitation stems from the geometric structure of the cortical surface, which suffers from poor global connectivity, creating bottlenecks that impede the flow of information Acknowledgments and Funding S.D.A.

Title and authors: Marcus: So we're moving into segment two where we look at the title and authors of this paper, "Graph Neural Network Reveals the Cortical Morphology of Local Brain Aging in Normal Cognition and Alzheimer’s Disease".

Ines: The title itself tells us that the focus is on using a graph neural network to see how the shape or morphology of the cortex relates to aging specifically for local brain areas.

Marcus: It’s not just looking at one big age number for the whole brain; it’s about mapping out where in your cortex things are changing as you get older, both when you're healthy and when you have Alzheimer’s Disease.

Yuki: And having this focus on local patterns is important because aging isn't uniform across the brain; it happens differently in different regions.

Ines: The authors are Anderson, Jomsky, Chaudhari, Chowdhury and Zheng. They come from a mix of computational biology and engineering backgrounds at institutions like USC and Berkeley.

Marcus: That background diversity is reflected in the methodology; they're using complex machine learning tools to analyze high-resolution MRI data from a large cohort of normal adults.

Yuki: From a population genetics perspective, having that many normal subjects helps them establish a baseline for what 'normal' cortical aging actually looks like before introducing disease variation.

Ines: They are essentially trying to build a framework where we can quantify biological age by looking at the observed variances of aging profiles, which is called brain age or BA.

Marcus: And they introduce the concept of the BAG gap, which is the difference between chronological age and brain age, reflecting advanced or delayed brain aging.

Yuki: That gap idea connects to broader population studies about how individual biological clocks might differ across a population.

Ines: So in short, this paper proposes a graph neural network that uses cortical features to estimate local brain age, which is much more detailed than just the global brain age they've studied before.

Marcus: It sets up the expectation for a study that can provide high-resolution data on aging patterns across different clinical states.

Yuki: I’m interested in how this helps us understand if these local changes are just noise, or if they actually reflect real biological processes related to neurodegeneration.

The paper's summary: Ines: Now let’s talk about the actual summary of the paper, "Graph Neural Network Reveals the Cortical Morphology of Local Brain Aging in Normal Cognition and Alzheimer’s Disease".

Marcus: The core idea is that they use a graph neural network to estimate local brain age (LBA) across the cortical surface at a high spatial resolution, with an average inter-vertex distance of one point three seven millimeters.

Yuki: So they are not just guessing a single brain age number; they are creating cortically specific predictions for every point on the surface.

Ines: They input morphometric features like cortical thickness surface area curvature GWR and sulcal depth to this network, and the goal is to estimate LBA across that entire cortical surface mesh.

Marcus: They trained this model on cortical meshes from fourteen thousand four hundred twenty-three cognitively normal adults from datasets including UKBB NACC and IXI <ref:2601.10912#pg2>.

Yuki: So the sample size is quite large, which gives them a good foundation for generalizing these findings to larger groups like ADNI.

Ines: The paper says that this approach provides a framework for quantifying anatomical brain aging, moving beyond just global BA to focus on what's happening at the subject level.

Marcus: They found that in cognitively normal adults, association cortices emerged as the primary sites for morphometric aging when looking at this LBA estimation.

Yuki: That means they’re identifying specific functional areas that are most sensitive to age changes during healthy aging.

Ines: And then they show how these patterns change when looking at the ADNI dataset, showing regional differences in LBAGs that are much larger in certain AD regions like the inferior temporal gyrus and parahippocampal gyrus.

Marcus: They also used feature ablation to pinpoint which inputs contribute most strongly to those local aging effects in specific areas.

Yuki: That’s a nice step because it shows that surface area might be more sensitive than other features in certain regions.

The paper's improvements: Ines: Let’s discuss what the paper suggests as improvements to their work and what those suggest for future research.

Marcus: They show that they can quantify how different morphometric properties capture distinct aspects of brain aging by calculating the difference in ablation effects between cohorts, which they call a positive Delta alpha ff indicating that the feature contributes more strongly to disease-specific aging.

Yuki: That’s a really sophisticated way to compare features across groups, moving beyond just seeing which one is 'best' overall.

Ines: They also found that surface area demonstrated its strongest negative effects throughout the visual and parietal association cortices when comparing normal aging versus AD.

Marcus: And they found that curvature and GWR showed their strongest effects in those temporal regions, particularly the inferior temporal gyrus and parahippocampal gyrus.

Yuki: It’s interesting because it shows that different features are important in different contexts, which is a bit more nuanced than just saying one feature is universally important.

Ines: The paper also suggests that the parahippocampal gyrus is implicated in both mild cognitive impairment and Alzheimer’s Disease, which points to it being a sensitive marker of disease progression.

Marcus: That's a strong finding because it suggests that region could be an early indicator for clinical changes.

Yuki: It supports the idea that this specific structure might be a good candidate for future biomarker research connecting morphology to clinical scores.

Conclusion: Ines: So to wrap up the paper, the "Graph Neural Network Reveals the Cortical Morphology of Local Brain Aging in Normal Cognition and Alzheimer’s Disease". The main implication is that we now have a way to estimate local brain age.

Marcus: This method allows us to see precisely where aging is happening on a subject's cortex, which gives us detailed anatomical information for both healthy aging and disease.

Yuki: And it moves the field toward using these local patterns as tools for tracking progression in conditions like Alzheimer’s Disease.

Ines: The paper shows that different morphometric features are important in different contexts, which is a key insight we need to keep in mind for future studies.

Marcus: It also suggests that surface area might be more sensitive to normal aging than to the decline seen in AD, which is a nuance worth focusing on.

Yuki: Ultimately, identifying regions like the parahippocampal gyrus as being involved across MCI and AD suggests it could be a good candidate for early disease detection efforts.

Ines: So this work provides a way to quantify local brain aging using sophisticated machine learning to see what it means for the world.