Abstract
Shape modelling describes methods aimed at capturing the natural variability of shapes and commonly relies on probabilistic interpretations of dimensionality reduction techniques such as principal component analysis. Due to their computational complexity when dealing with dense deformation models such as diffeomorphisms, previous attempts have focused on explicitly reducing their dimension, diminishing de facto their flexibility and ability to model complex shapes such as brains. In this paper, we present a generative model of shape that allows the covariance structure of deformations to be captured without squashing their domain, resulting in better normalisation. An efficient inference scheme based on Gauss-Newton optimisation is used, which enables processing of 3D neuroimaging data. We trained this algorithm on segmented brains from the OASIS database, generating physiologically meaningful deformation trajectories. To prove the model's robustness, we applied it to unseen data, which resulted in equivalent fitting scores.
Introduction
In neuroimaging, "normalization" is the critical process of warping individual brain scans into a common coordinate space to facilitate population-level comparisons. However, human neuroanatomy is highly variable, and capturing this variability requires sophisticated models of deformation. Ideally, these models should account for both "typical" population trends (e.g., age-related atrophy) and individual "anatomical noise."
Modeling these deformations—specifically diffeomorphisms, which are smooth and invertible transformations—is computationally intensive. Traditional approaches often rely on dimensionality reduction like Principal Component Analysis (PCA) to make the problem tractable. However, standard PCA-based methods often "squash" the data too much, discarding fine-grained details to maintain computational feasibility. This results in models that are too rigid to capture the true complexity of human neuroanatomy. This paper introduces a generative model that captures population-wide variability while maintaining the flexibility to account for individual differences.
Methodology
The proposed generative shape model is based on a Probabilistic PCA (PPCA) framework with several key innovations designed for efficiency and flexibility.
The Generative Model
The model defines the initial velocity of a deformation ($v_n$) as a combination of a principal subspace and a residual field: $$v_n = Wz_n + r_n$$ Here, $Wz_n$ represents the major trends in the population (the principal components), while $r_n$ captures the "anatomical noise" unique to an individual. By explicitly modeling the residual field, the authors ensure the model doesn't force every brain into a low-dimensional "box," allowing for superior alignment compared to rigid PCA models.
Inference and Optimization
To handle the non-linearity of warping images, the authors utilize a Gauss-Newton (GN) optimization scheme. GN is more efficient than standard gradient descent because it leverages second-order information (the Hessian) to find optimal shape coordinates faster and more reliably. Furthermore, the model operates on segmented tissue probability maps (grey and white matter) rather than raw MRI intensities, making the registration process more robust to noise and artifacts.
Experimental Setup
The model was trained and evaluated using the OASIS (Open Access Series of Imaging Studies) database. The training set consisted of 38 subjects, with an additional 38 unseen subjects used for testing robustness. Images were pre-segmented using FSL and smoothed with a 1-voxel Gaussian kernel to create the tissue probability maps used by the model.
Results and Discussion
Qualitative Improvement and Detail
As the number of principal modes in the model increased, the reconstructed brain template became significantly sharper and more detailed. The inclusion of the residual field ("+R") further enhanced the model's ability to capture fine-grained anatomical features.
Robustness and Biological Validation
The model demonstrated excellent generalization, performing almost identically on seen versus unseen data. The categorical log-likelihood for the training set was $-7.23 \pm 1.26$, while the test set was $-7.55 \pm 0.88$. This minimal difference indicates that the model is learning generalizable anatomical features rather than over-fitting to the training set.
Crucially, the model's first principal component (PG1) was found to correlate strongly with subject age. This suggests that the mathematical framework successfully captures real biological processes—such as the typical patterns of brain aging—without being explicitly supervised to do so. Additionally, latent space analysis successfully mapped trajectories associated with Alzheimer’s Disease, hinting at the model's potential as a diagnostic tool.
Conclusion
This work presents a robust, efficient Bayesian framework for diffeomorphic shape modeling. By combining Gauss-Newton optimization with a dual-component velocity field (principal modes plus residuals), the authors achieve high-fidelity brain normalization that respects biological complexity. Future work will focus on capturing sparse, local covariance patterns and integrating the shape model directly with segmentation algorithms to process raw MRI data.
Summary updated