Learning a Probabilistic Model for Diffeomorphic Registration

Julian Krebs, Hervé Delingette, Boris Mailhé, Nicholas Ayache, Tommaso Mansi

TL;DR

This paper introduces an unsupervised deep learning framework that learns a low-dimensional, probabilistic "map" of how images deform. By combining a Conditional Variational Autoencoder (CVAE) with diffeomorphic constraints, the model achieves state-of-the-art 3D cardiac registration in a fraction of a second while enabling advanced tasks like disease clustering and "deformation transport" between different patients.

What Problem Does This Paper Address?

Medical image registration—the process of finding voxel-to-voxel correspondences between two images—is a cornerstone of clinical analysis. Traditionally, this is solved via slow, iterative numerical optimization. While recent Deep Learning (DL) methods have sped this up, they face several hurdles:

  1. Supervision Gap: It is nearly impossible to get "ground truth" voxel mappings for training.
  2. Lack of Structure: Most unsupervised DL methods produce a deformation field but don't provide a structured way to analyze or compare those deformations across a population.
  3. Physical Realism: Many models produce "folding" in the deformation fields, which is physically impossible in biological tissue.

The authors aim to create a model that is not only fast and unsupervised but also learns a structured, low-dimensional latent space where similar deformations (e.g., the specific contraction pattern of a diseased heart) are grouped together.

Methods

The authors propose a Conditional Variational Autoencoder (CVAE) architecture:

  • Unsupervised Learning: The network learns by minimizing a similarity loss (Symmetric Local Cross-Correlation) between a warped moving image and a fixed target image.
  • Probabilistic Latent Space: The encoder maps the difference between two images into a low-dimensional vector ($z$). The decoder is "conditioned" on the moving image, forcing the latent vector to encode only the motion rather than the image's appearance.
  • Diffeomorphic Constraints: To ensure the deformations are smooth and invertible (no folding), the model predicts a Stationary Velocity Field (SVF). This field is passed through a differentiable "exponentiation layer" (using the scaling and squaring algorithm) to produce the final transformation.
  • Multi-Scale Architecture: The model estimates deformations at three different resolutions, which helps capture both large-scale global motion and fine-grained local details.

Datasets

The model was evaluated on 3D intra-subject registration using 334 cardiac cine-MRI sequences.

  • Training: 234 cases.
  • Testing: 100 cases from the ACDC (Automatic Cardiac Diagnosis Challenge) at STACOM 2017.
  • Diversity: The dataset includes healthy subjects and patients with four types of cardiac disease: dilated cardiomyopathy (DCM), hypertrophic cardiomyopathy (HCM), previous myocardial infarction (MINF), and abnormal right ventricle (RV).

Key Insights and Ideas

  • Deformation as a Language: By forcing the model to compress motion into a 32-dimensional vector, the authors essentially created a "feature space" for anatomy in motion.
  • Decoupling Appearance from Motion: By feeding the moving image directly into the decoder, the latent space is "cleaned" of anatomical appearance details, focusing purely on the geometric transformation.
  • Latent Space Clinical Utility: The latent space isn't just a mathematical convenience; it has clinical meaning. A simple linear projection of these 32D vectors could classify the five cardiac disease states with 83% accuracy, suggesting the model "discovered" the mechanical signatures of these diseases.
  • Deformation Transport: Because the latent space is standardized, you can take the "contraction signature" ($z$-code) of a diseased heart and apply it to a healthy heart's geometry to simulate how that specific pathology would look in a different patient.

Key Results

  • Accuracy: Achieved a mean DICE score of 81.2% and a mean Hausdorff distance of 7.3mm, outperforming or matching state-of-the-art methods like SyN and VoxelMorph.
  • Speed: Registration takes only 0.32 seconds on a modern GPU, compared to minutes for traditional iterative methods.
  • Regularity: The deformations produced were significantly smoother and more "regular" than other DL-based approaches, as measured by the gradient of the Jacobian determinant.
  • Generative Power: The model can sample from the latent space to generate an infinite variety of realistic, synthetic deformations for data augmentation or simulation.

Limitations and Future Work

  • Prior Assumptions: The model assumes a simple multivariate Gaussian distribution for the latent space. The authors suggest that a more complex "mixture of Gaussians" might better represent the distinct clusters of different diseases.
  • Uncertainty Quantification: While the probabilistic nature of the VAE allows for estimating registration uncertainty (by sampling multiple $z$ vectors), the authors did not explore this clinical application in depth in this specific paper.
  • Latent Dimensionality: The choice of 32 dimensions was a heuristic trade-off; finding the "optimal" bottleneck size for different organs or diseases remains an open question.

Summary updated

Abstract

We propose to learn a low-dimensional probabilistic deformation model from data which can be used for the registration and the analysis of deformations. The latent variable model maps similar deformations close to each other in an encoding space. It enables to compare deformations, to generate normal or pathological deformations for any new image, or to transport deformations from one image pair to any other image. Our unsupervised method is based on the variational inference. In particular, we use a conditional variational autoencoder network and constrain transformations to be symmetric and diffeomorphic by applying a differentiable exponentiation layer with a symmetric loss function. We also present a formulation that includes spatial regularization such as the diffusion-based filters. In addition, our framework provides multi-scale velocity field estimations. We evaluated our method on 3-D intra-subject registration using 334 cardiac cine-MRIs. On this dataset, our method showed the state-of-the-art performance with a mean DICE score of 81.2% and a mean Hausdorff distance of 7.3 mm using 32 latent dimensions compared to three state-of-the-art methods while also demonstrating more regular deformation fields. The average time per registration was 0.32 s. Besides, we visualized the learned latent space and showed that the encoded deformations can be used to transport deformations and to cluster diseases with a classification accuracy of 83% after applying a linear projection.
BibTeX
@article{krebs2019learning,
  author = {Julian Krebs and Hervé Delingette and Boris Mailhé and Nicholas Ayache and Tommaso Mansi},
  title = {{Learning a Probabilistic Model for Diffeomorphic Registration}},
  journal = {IEEE Trans. Medical Imaging},
  year = {2019},
  url = {https://openreview.net/forum?id=ESW1gpZ7ON},
}

Loading related content...

Figure