Introduces gyral folding-based brain networks using three-hinge gyri to distinguish Alzheimer's from Lewy body dementia, addressing atlas inconsistencies. Applies AI network analysis to personalized neuroimaging.
The study introduces PaIRWaL, a probability-invariant random-walk learning framework that utilizes gyral folding-based cortical similarity networks to classify Alzheimer’s disease (AD) and Lewy body dementia (LBD) by addressing individual anatomical variability that often undermines traditional atlas-based brain network models . By using three-hinge gyri (3HGs) as biologically grounded, individualized anatomical landmarks, the method constructs cortical networks that reflect personalized folding patterns, overcoming the limitations of fixed parcellation schemes which may obscure patient-specific structural details .
PaIRWaL models brain networks through distributions of anonymized random walks rather than relying on explicit node correspondences, enabling permutation invariance and robustness to variable graph sizes across subjects . This approach naturally handles inter-individual differences in cortical folding and pathological heterogeneity, which commonly disrupt node alignment in conventional graph learning methods. The framework incorporates an Anatomy-Aware Anonymized Walk Recording (A³WR) module that encodes structural transitions and neighborhood relations while integrating anatomical priors via region-of-interest (ROI) attribute tokens, preserving both biological interpretability and invariance .
Cortical similarity networks are built from local morphometric features—such as cortical thickness, curvature, sulcal depth, surface area, and gray matter volume—extracted within 5-ring neighborhoods of each 3HG . Inter-node similarity is computed using symmetric Kullback-Leibler divergence with multivariate kernel density estimation, and graphs are constructed to ensure connectivity by combining a minimum spanning tree with the top 10% of strongest similarity edges .
Evaluated on a clinical cohort of 303 individuals (108 cognitively normal controls, 90 AD, and 105 LBD) from the University of Cambridge, including data from the MILOS project, PaIRWaL demonstrated consistent improvements over both gyral folding-based and atlas-based classification models . It achieved superior performance in differentiating AD from LBD—a challenging task due to overlapping clinical features—with significant gains in sensitivity and specificity (p < 0.05) . Ablation studies confirmed the importance of anonymization, non-backtracking, and anatomical tokens, with the largest performance drops observed when these components were removed .
Sensitivity analysis revealed optimal performance with random walk lengths between 64 and 128 and as few as 8 sampled walks per graph, ensuring computational efficiency . Visualization of aggregated 3HG heatmaps highlighted spatially localized, disease-relevant folding regions, suggesting the model captures neurobiologically meaningful structural alterations . These results underscore PaIRWaL’s potential as a reliable, interpretable AI-driven tool for differential dementia diagnosis using personalized neuroimaging data
This research addresses the clinical and computational challenge of differentiating between Alzheimer’s Disease (AD) and Lewy Body Dementia (LBD), two conditions with overlapping symptoms but distinct pathologies. To overcome the limitations of traditional atlas-based brain parcellation—which often suffer from registration inconsistencies and lack of anatomical precision—the authors propose a framework utilizing gyral folding-based cortical similarity networks. Specifically, the method leverages "three-hinge gyri," fundamental anatomical landmarks in the brain's folding structure, as nodes to construct a personalized connectivity map. This approach shifts the analysis from generalized region-of-interest templates to a subject-specific, geometry-driven representation of cortical architecture, allowing for a more granular examination of neurodegenerative changes.
The core technical contribution of the paper is the introduction of a Probability-Invariant Random Walk (PIRW) learning algorithm applied to these structural networks. By employing random walk strategies, the model captures deep topological relationships and structural connectivity patterns between gyral nodes, effectively learning node embeddings that reflect the brain's underlying geometry. The "probability-invariant" nature of the algorithm is designed to handle noise and variability in neuroimaging data, ensuring that the learned features are robust and consistent across different subjects. This enables the AI to identify subtle, localized morphological signatures associated with specific dementias that are often obscured in whole-brain or atlas-dependent analyses.
This work is significant because it bridges the gap between structural neuroanatomy and advanced graph machine learning, offering a robust tool for differential diagnosis. By decoupling the analysis from standardized atlases, the method reduces errors introduced by spatial normalization and provides a more faithful representation of an individual patient's neurobiology. Consequently, this approach holds promise for improving diagnostic accuracy in clinical settings, potentially enabling earlier and more reliable distinction between AD and LBD, which is crucial for determining appropriate treatment plans and patient management.
This study introduces a novel framework for distinguishing between Alzheimer’s disease (AD) and Lewy body dementia (LBD) by leveraging gyral folding-based cortical similarity networks. The approach addresses a critical challenge in neuroimaging: the inconsistency of traditional brain atlases in capturing individual anatomical variations. By constructing cortical networks using three-hinge gyri—a method that captures local folding patterns—researchers generate personalized, structurally informed representations of brain connectivity. These networks are then analyzed using probability-invariant random walk learning, a technique that enhances robustness to noise and atlas variability while preserving meaningful topological features.
The key contributions of this work include: 1. A gyral folding-based network construction method that improves on atlas-dependent approaches by focusing on intrinsic cortical geometry. 2. AI-driven network analysis (via random walk learning) to extract discriminative features for AD vs. LBD classification, demonstrating superior performance in distinguishing these dementias. 3. Addressing atlas inconsistencies, which are a major limitation in neuroimaging-based diagnostics.
This research matters because it advances personalized neuroimaging diagnostics by moving beyond rigid atlas-based methods toward individualized brain network representations. The ability to reliably differentiate AD and LBD—two conditions with overlapping symptoms but distinct pathological mechanisms—has significant clinical implications for early and accurate diagnosis. By improving the specificity of neuroimaging markers, this approach could enhance treatment stratification and monitoring in dementia research.
Source: [arXiv:2602.17557](https://arxiv.org/abs/2602.17557)