Probabilistic non-linear registration with spatially adaptive regularisation

Area of Science:

• Medical Imaging Analysis
• Computational Anatomy
• Neuroimaging

Background:

• Non-linear registration is crucial for medical image analysis, but global regularization can limit accuracy.
• Spatially varying regularization is needed to adapt to complex anatomical changes.

Purpose of the Study:

• To develop a novel method for inferring spatially varying regularization in non-linear registration using Bayesian inference.
• To improve the flexibility and data-driven nature of regularization in image registration.

Main Methods:

• Full Bayesian inference on a probabilistic registration model.
• Parameterization of the transformation prior using a weighted mixture of spatially localized components.
• Adaptive determination of prior influence based on local data support.

Main Results:

• The proposed method allows for reduced prior influence in data-rich areas and stronger constraints in less informative regions.
• Spatially adaptive priors lead to sparser deformations and better localization of regional volume changes.
• Results show more data-driven and localized maps of registration uncertainty.

Conclusions:

• The novel spatially adaptive prior reduces unwanted impacts of regularization on inferred transformations.
• This method is particularly beneficial for applications like tensor-based morphometry, aiding in the analysis of diseases such as Alzheimer's.
• Demonstrates the first use of Bayesian model comparison for selecting regularization types in this context.