Abstract
This paper introduces high-speed approximation schemes for geodesic regression to model 3D shape evolution. While geodesic regression is mathematically rigorous, its computational cost often limits clinical application. By applying a novel hierarchy of regression models—including a hybrid Geodesic Regression with Linear Residuals (GRLR)—to longitudinal brain scans, the authors characterize subtle "shear" deformations in the female hippocampus during progesterone fluctuations. These findings demonstrate that shape analysis can capture anatomical changes missed by traditional volumetric metrics, providing a new biomarker for hormonal influences on brain anatomy.
Introduction
Neuroimaging research focusing on female-specific hormonal transitions remains significantly underrepresented, despite women's higher risk for conditions like Alzheimer’s disease. The hippocampus, a primary site for early Alzheimer's pathology, is known to undergo sub-regional changes during the menstrual cycle. However, total hippocampal volume often remains stable, suggesting that volume is an insufficient metric for capturing these dynamics. This study proposes using shape analysis, specifically geodesic regression, to treat 3D surfaces as points on a curved manifold. To overcome the computational bottlenecks of standard geodesic regression, the authors develop efficient approximation schemes.
Methods
The authors propose a hierarchy of regression models to balance speed and accuracy:
- Linear Regression (LR): The fastest baseline, treating 3D coordinates as flat Euclidean space. While efficient, it lacks parameterization invariance.
- Geodesic Regression (GR): The "gold standard" that respects the intrinsic geometry of the shape space using a second-order Sobolev metric. It is computationally intensive due to the need for numerical integration.
- Geodesic Regression with Linear Residuals (GRLR): A novel hybrid approach. It uses geodesic modeling to describe shape change but calculates residuals using simple Euclidean distance, avoiding the expensive "logarithm map."
To guide model selection, the authors introduce two "Rules of Thumb":
- $\Delta$-Test: If the total change in the dataset is small compared to the curvature of the shape space, use Linear Regression.
- $\delta$-Test: If the measurement noise is small compared to the curvature, use GRLR.
These methods were validated on synthetic ellipsoids and a "dense-sampling" longitudinal dataset consisting of 11 high-resolution MRIs from a single female subject, captured daily during the progesterone peak of her menstrual cycle.
Results
Synthetic tests demonstrated that Linear Regression can achieve speedups of up to 14 million times ($14 \times 10^6$) compared to full GR, with an error rate below 0.05% for small deformations. The GRLR hybrid significantly outperformed standard GR in speed while maintaining high accuracy, making it feasible for standard hardware.
Applying the model to the 11-day MRI sequence revealed that an increase in progesterone corresponds specifically to a shear deformation (sliding/tilting) of the hippocampal formation. This insight is significant because the structure changes its functional geometry without a corresponding change in total volume, providing a new biomarker for how sex hormones influence brain anatomy.
Discussion and Conclusion
The study highlights that shape analysis can reveal subtle anatomical changes that traditional volumetric analyses miss. By identifying the "Log map" as the primary computational bottleneck, the authors created the GRLR hybrid, offering a "best of both worlds" scenario for biological applications.
While the real-world application was limited to a single-subject dense-sampling study, the results provide a powerful proof of concept. Future work will focus on validating these findings across larger, more diverse populations and determining if these subtle shape changes can predict the onset of neurological diseases earlier than volume-based metrics. The implementation has been made publicly available to encourage further research in biomedicine and computer vision.
Summary updated