Gaussian process interpolation for uncertainty estimation in image registration.

Area of Science:

• Medical imaging
• Computational anatomy
• Machine learning

Background:

• Intensity-based image registration necessitates resampling images onto a common grid for similarity evaluation.
• Interpolation uncertainty varies spatially based on resampled point locations relative to the grid.

Purpose of the Study:

• To develop a novel image registration method that quantifies and incorporates interpolation uncertainty.
• To improve the accuracy of intensity-based image registration by accounting for spatial uncertainty.

Main Methods:

• Utilizing Bayesian inference with Gaussian processes to model image distributions.
• Employing the Gaussian process posterior covariance matrix to estimate interpolation uncertainty.
• Developing a generative model for registration by integrating Gaussian processes.

Main Results:

• A new similarity measure was derived by marginalizing over resampled images, incorporating interpolation uncertainty.
• The proposed method demonstrated increased registration accuracy compared to existing approaches.
• An efficient approximation scheme was developed for seamless integration with current registration techniques.

Conclusions:

• Bayesian inference with Gaussian processes offers a robust framework for uncertainty-aware image registration.
• Accounting for interpolation uncertainty can significantly improve registration performance.
• The developed approximation scheme facilitates practical application in various medical imaging scenarios.