Learning a Probabilistic Model for Diffeomorphic Registration.

Area of Science:

• Medical imaging analysis
• Computational anatomy
• Machine learning for healthcare

Background:

• Accurate registration and analysis of deformations are crucial for medical image interpretation.
• Existing methods often lack the ability to generate realistic deformations or compare complex deformation patterns.

Purpose of the Study:

• To develop a low-dimensional probabilistic model for learning, analyzing, and generating deformations from data.
• To enable comparison, generation, and transportation of deformations across different medical images.
• To achieve state-of-the-art performance in 3D intra-subject registration of cardiac cine-MRIs.

Main Methods:

• Utilizing unsupervised learning with variational inference and a conditional variational autoencoder network.
• Ensuring symmetric and diffeomorphic transformations via a differentiable exponentiation layer and symmetric loss function.
• Incorporating spatial regularization (e.g., diffusion-based filters) and multi-scale velocity field estimations.

Main Results:

• Achieved state-of-the-art registration performance on 3D cardiac cine-MRIs (mean DICE score 81.2%, mean Hausdorff distance 7.3 mm).
• Demonstrated more regular deformation fields compared to existing methods, with rapid registration times (0.32 s).
• Successfully used the learned latent space for deformation transportation and disease clustering (83% classification accuracy).

Conclusions:

• The proposed probabilistic deformation model offers a powerful tool for medical image registration and analysis.
• The method's ability to generate and transport deformations has implications for data augmentation and understanding disease progression.
• This framework advances unsupervised learning applications in medical imaging, providing efficient and accurate deformation modeling.