Spatially varying registration using Gaussian processes.

Area of Science:

• Medical Imaging
• Computational Anatomy
• Machine Learning

Background:

• Accurate medical image registration is crucial for diagnosis and treatment planning.
• Existing registration methods often struggle with spatially varying image noise and artifacts.
• Spatially-varying registration requires adaptable smoothness constraints.

Purpose of the Study:

• To develop a novel approach for spatially-varying registration using Gaussian process priors.
• To enhance robustness against noise in medical image datasets, such as cone beam CT scans.
• To provide a flexible method applicable to various registration algorithms.

Main Methods:

• Implementation of spectral tempering to modify Gaussian process spectra.
• Creation of non-stationary Gaussian processes with user-defined tempering functions.
• Integration of the modified prior into existing registration frameworks without altering the core algorithm.

Main Results:

• Demonstrated robustness against noise in atlas-based skull registration of cone beam CT images.
• Achieved accurate correspondence in noise-free regions of the datasets.
• Showcased the method's applicability to spline-based models and statistical shape/deformation models.

Conclusions:

• The proposed Gaussian process prior with spectral tempering offers an effective solution for spatially-varying registration.
• This method significantly improves registration performance in the presence of noise, a common challenge in medical imaging.
• The approach is versatile, allowing for spatially-varying adaptations of various registration techniques.