Abstract
Shape analysis is critical for the early diagnosis and prognosis of neurodegenerative diseases such as Alzheimer’s Disease (AD). However, obtaining effective and robust shape representations for the complex, folded geometry of the human cortex remains a significant challenge. This paper introduces the use of Alexandrov polyhedra as intrinsic, surface-based shape signatures for cortical morphometry. By reformulating the classical Minkowski problem as a novel spherical optimal transport (OT) task, we develop an efficient algorithm to represent closed genus-0 surfaces as unique convex polyhedra. Our experimental results demonstrate that Alexandrov polyhedra can distinguish between pathology-confirmed AD patients and cognitively unimpaired (CU) individuals with higher sensitivity than traditional biomarkers. Furthermore, we provide a visualization method to pinpoint regional structural changes, such as those in the temporal lobe, impacted by AD.
Introduction
Early detection of Alzheimer’s Disease often depends on identifying subtle morphological changes in brain structure. While structural MRI (sMRI) is widely used, clinical analysis frequently relies on coarse metrics like volume or surface area. Surface-based analysis offers a more granular view but is complicated by the "wrinkled" nature of the cortex. Existing methods like conformal mapping can suffer from extreme area distortion, while other optimal transport techniques often rely on projections that introduce artificial biases.
This research proposes an intrinsic approach. By representing the cortex as an Alexandrov polyhedron, we capture its Riemannian metric in a convex form that is mathematically rigorous and computationally efficient.
Methodology
The core of our approach lies in solving the Minkowski problem: reconstructing a convex shape from its Gaussian curvature. For a genus-0 surface, there exists a unique convex representation—the Alexandrov polyhedron.
Spherical Optimal Transport
We reformulate the reconstruction as an optimal transport problem on a sphere. We aim to transport the area elements of the brain surface to a target measure on a unit sphere in the most "economical" way. Unlike methods that project to a 2D plane, our method operates directly on the sphere using spherical metrics, ensuring independence from initial conditions and avoiding artificial distortions.
Algorithmic Innovations
To make this practical for large-scale clinical use, we introduced several optimizations:
- Variational Formulation: We optimize a convex functional energy using a novel formulation of the gradient and Hessian, enabling fast convergence via Newton’s method.
- Localized Edge Flips: To manage mesh connectivity changes during optimization, we implemented a generalized Lawson’s edge flip technique. This avoids the need to rebuild the entire mesh at each iteration, significantly boosting speed.
Experimental Results
We validated our method using data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI), comprising 80 AD patients (Aβ+) and 99 cognitively unimpaired (CU) controls (Aβ-).
Sensitivity and Performance
The proposed Alexandrov Polyhedron Hausdorff Distance (APHD) demonstrated superior sensitivity in group difference tests. While traditional biomarkers like Volume ($p=0.703$) and Surface Area ($p=0.486$) failed to reach statistical significance, APHD achieved a significant p-value of 0.034.
Computationally, our algorithm averaged 16.11 seconds per surface, representing a 27% speed improvement over previous state-of-the-art intrinsic OT methods.
Anatomical Localization
Our visualization method compares local geometry differences across surfaces. The resulting heat maps successfully identified significant structural changes in the temporal lobe and posterior cingulate, regions known to be early targets of Alzheimer's pathology.
Conclusion
Alexandrov polyhedra provide a powerful and sensitive "geometric signature" for neurodegenerative disease analysis. By capturing intrinsic geometric changes that traditional metrics miss, this method offers a promising new tool for clinical research. Future work will explore its application to other conditions, such as Parkinson's disease and schizophrenia, to evaluate its potential as a universal geometric biomarker.
Summary updated