Abstract
We describe a diffeomorphic registration algorithm that allows groups of images to be accurately aligned to a common space, intended for incorporation into the SPM software. By performing inference in a probabilistic graphical model that accounts for variability in both shape and appearance, we provide a general and entirely unsupervised framework. The model assumes individual anatomies are generated by deforming a latent 'average' brain. Agnostic to imaging modality and requiring no prior processing, the algorithm achieves state-of-the-art results on manually labelled datasets, including a 17% increase in overlap score on unprocessed data compared to existing methods.
Introduction
Image registration—the alignment of multiple 3D scans into a common coordinate space—is fundamental to neuroimaging. Traditional tools often require extensive pre-processing, such as skull-stripping or intensity normalization, and are frequently limited to specific MRI contrasts. Furthermore, registration presents a "chicken and egg" problem: defining a standard brain template requires well-aligned images, yet high-quality alignment requires a reliable template.
The proposed MultiBrain framework addresses these challenges by treating registration as a probabilistic inference problem within a generative model. This unsupervised approach allows for high-quality registration of raw data from diverse scanners and modalities without manual intervention.
Methods: A Generative Bayesian Model
The core of the framework is a joint probability distribution describing how a brain scan is generated from two latent components: Shape ($S$) and Appearance ($A$).
- Shape ($S$): A latent tissue template (the "average" brain) undergoes a subject-specific diffeomorphic transformation. This transformation is smooth, invertible, and combines rigid global alignment with complex nonlinear warping via geodesic shooting.
- Appearance ($A$): Subject-specific Gaussian Mixture Models (GMMs) account for varying tissue intensities. The model explicitly incorporates Intensity Non-Uniformity (INU) to handle magnetic field bias artifacts common in MRI.
The model is solved using a Variational Expectation-Maximization (VEM) algorithm. This iterative process simultaneously refines the estimate of the "average brain" and the individual subject warps. To ensure stability and avoid local minima, the framework employs multi-scale optimization and the Böhning bound for Hessian approximation.
Experimental Results
The algorithm was evaluated on three benchmarks: the "cleaned" LPBA40 dataset, the unprocessed IBSR18 dataset, and a large-scale pre-learned model (MB-L) trained on 277 scans.
Key Findings
- Superior Accuracy: MultiBrain achieved a 17% increase in overlap score on the challenging, unprocessed IBSR18 dataset compared to baseline methods.
- Benchmarking: The framework outperformed established gold standards like ANTs (SyN) and ART. It also showed superior robustness compared to deep learning models like VoxelMorph, which often struggle with unprocessed data.
- Modality Agnosticism: By modeling the bias field and noise directly, the algorithm learns appearance on the fly, making it theoretically applicable to various MRI contrasts (T1, T2, etc.) without modification.
Discussion
MultiBrain demonstrates that modeling the underlying imaging process—rather than just matching raw pixel intensities—leads to more robust registration. While the current CPU-based implementation takes 15–30 minutes per scan, it provides a level of accuracy and flexibility that is highly valuable for clinical and research workflows. Future work will focus on GPU acceleration and extending the model to automated segmentation and semi-supervised learning.
Summary updated