Abstract
In this paper we propose a novel framework to unite a population to an optimal (unknown) pose through their mutual deformation. The registration criterion comprises three terms, the first imposes compactness on appearance of the registered population at the pixel level, the second tries to minimize the individual distances between all possible pairs of images, while the last is a regularization one imposing smoothness on the deformation fields. The problem is reformulated as a graphical model that consists of hidden (deformation fields) and observed variables (intensities). A novel deformation grid-based scheme is proposed that guarantees the diffeomorphism of the deformation and is computationally favorably compared to standard deformation methods. Towards addressing important deformations we propose a compositional approach where the deformations are recovered through the sub-optimal solutions of successive discrete MRFs by using efficient linear programming. Promising experimental results using real 2D data demonstrate the potentials of our approach.
Introduction
A fundamental challenge in medical image analysis is the alignment of a population of images into a common coordinate space. Traditional methods often rely on a pre-defined reference template, which introduces "reference bias"—the final alignment depends heavily on the choice of the initial template. While group-wise registration avoids this by aligning all images simultaneously to an unknown optimal pose, it is historically computationally intensive and prone to local minima. This paper introduces a novel graphical model framework that achieves unbiased, diffeomorphic group-wise registration efficiently.
Methodology
The Graphical Model Framework
The authors reformulate population registration as a discrete labeling problem using a Markov Random Field (MRF). Instead of continuous warping, a grid of control points is superimposed over the images. The objective function integrates three critical components:
- Global Compactness: Measures the statistical consistency (e.g., entropy) of the entire image stack.
- Local Pairwise Relations: Minimizes distances between all image pairs to ensure local consistency.
- Smoothness Regularization: Prevents non-physical deformations like "tearing" or "folding."
Deformation Model and Optimization
To ensure computational efficiency, the framework utilizes Hermite splines, which offer smoothness comparable to cubic B-splines but with lower overhead. A key feature is the guaranteed diffeomorphism: by constraining grid point movement to less than 0.25 times the grid spacing, the transformation remains smooth and invertible. The resulting MRF is solved using the Fast-PD algorithm, a robust discrete optimization technique based on linear programming.
Experimental Results
The method was validated on 2D MR images of human calf muscles. Accuracy was measured using Dice coefficients derived from expert segmentations.
Key findings include:
- Visual Clarity: The mean image of the aligned population is significantly sharper than those produced by traditional methods, indicating precise anatomical alignment.
- Quantitative Performance: The group-wise approach consistently outperformed state-of-the-art pairwise registration methods.
- Stability: The method showed low dispersion in Dice scores, suggesting high reliability across different input datasets.
- Efficiency: Registration of 20 images was completed in approximately 30 minutes using a Matlab implementation, with significant potential for speedup in C++.
Conclusion and Future Work
This framework provides a modular and bias-free solution for population registration. By treating the optimal pose as a hidden variable in a graphical model, it finds the natural "center" of a dataset. While currently demonstrated on 2D slices, the mathematical foundation supports extension to 3D volumes. Future work will explore higher-order MRFs for improved global cost approximation and multi-modal registration (e.g., MRI to CT).
Summary updated