Abstract
Magnetic resonance imaging (MRI)-derived hippocampus measurements have been associated with different cognitive domains. The knowledge of hippocampal structural deformations as we age has contributed to our understanding of the overall aging process. Different morphological hippocampal shape analysis methods have been developed, but it is unclear how their principles relate and how consistent are the published results in relation to cognition in the normal elderly in the light of the new deep-learning-based (DL) state-of-the-art modeling methods. We compared results from analysing the hippocampal morphology using manually-generated binary masks and a Laplacian-based deformation shape analysis method, with those resulting from analysing SynthSeg-generated hippocampal binary masks using a DL method based on the PointNet architecture, in relation to different cognitive domains. Whilst most previously reported statistically significant associations were also replicated, differences were also observed due to 1) differences in the binary masks and 2) differences in sensitivity between the methods. Differences in the template mesh, number of vertices of the template mesh, and their distribution did not impact the results.
Introduction
The hippocampus is a critical brain structure for memory and a primary target for aging and Alzheimer’s research. While volume is a common metric for assessing hippocampal health, it is often a blunt instrument for capturing subtle structural changes. Shape analysis—examining specific structural deformations like "dents" and "bulges"—provides richer data on exactly where atrophy occurs. However, traditional shape modeling techniques, such as Laplacian-based deformation, often "blur" details due to mathematical smoothing constraints. This study introduces a new deep-learning-based 3D shape modeling method designed to map hippocampal morphology with higher sensitivity, revealing how specific structural variations correlate with cognitive domains like memory and processing speed in older adults.
Methods
The researchers developed a three-stage AI-based hippocampal deformation framework to improve upon traditional geometric techniques:
- Target Individual Shape Extraction: Hippocampal "point clouds" are generated from MRI scans. This study compared manual segmentation with SynthSeg, an AI-based automated tool, to evaluate the impact of segmentation accuracy.
- AI-Based Iterative Shape Modelling: Instead of traditional differential geometry, the framework utilizes a Multi-Layer Perceptron (MLP) based on the PointNet architecture. Following the Point2Mesh approach, the model iteratively deforms a template hippocampal mesh to fit the individual's point cloud.
- Local Deformity Computation: By establishing vertex correspondence across all subjects, the "signed distance" of every point from an average template is calculated to quantify local deformations.
Results
The AI-driven pipeline was applied to the Lothian Birth Cohort 1936 (LBC1936), involving 601 participants (mean age ~72.6 years). Cognitive performance was assessed across three latent variables: g-memory, g-speed, and general fluid intelligence (g).
The AI method demonstrated significantly higher sensitivity than traditional Laplacian methods, capturing associations between hippocampal shape and cognition that were previously obscured by smoothing algorithms. A key finding was the importance of the initial segmentation; SynthSeg-generated masks were more consistent and yielded stronger statistical associations than manual masks, achieving a Dice Similarity Coefficient (DSC) of 0.978.
Notably, the method proved robust to variations in mesh density and vertex distribution. Doubling the number of vertices had negligible impact on clinical findings, suggesting the AI model effectively captures the underlying biological signal regardless of resolution. Anatomically, the AI method highlighted inward deformation in the CA1 subfield and hippocampal head as strongly linked to poorer memory, findings that align with high-resolution anatomical studies but are often missed by lower-sensitivity tools.
Discussion and Conclusion
While the AI method offers superior sensitivity, it remains sensitive to noise in the initial segmentation. If the initial MRI is noisy, the AI will faithfully reconstruct those errors. Future work will focus on applying this framework to longitudinal data to determine if shape change over time is a more potent predictor of cognitive decline than cross-sectional snapshots. Additionally, testing the model on more diverse datasets will be crucial for generalizing these findings across different populations and disease states.
Summary updated