Image registration: Maximum likelihood, minimum entropy and deep learning

Area of Science:

• Medical Imaging
• Computer Vision
• Computational Anatomy

Background:

• Image registration is crucial for analyzing medical data.
• Existing methods like mutual information have limitations.
• Groupwise registration requires robust theoretical frameworks.

Purpose of the Study:

• To develop a novel theoretical framework for image registration using maximum profile likelihood.
• To derive and optimize methods for pairwise and groupwise registration.
• To introduce a deep learning approach for enhanced registration accuracy.

Main Methods:

• Maximum profile likelihood framework for pairwise and groupwise registration.
• Asymptotic analysis to demonstrate entropy minimization.
• Derivation of the congealing method via profile likelihood optimization.
• Feature-based registration and deep metric registration using discriminative classifiers.

Main Results:

• Maximum profile likelihood registration minimizes an upper bound on joint entropy.
• The congealing method is derived for groupwise registration.
• A deep metric registration approach is proposed, outperforming standard methods on challenging datasets.
• The framework accommodates feature-based and deep learning registration strategies.

Conclusions:

• Maximum profile likelihood provides a powerful theoretical foundation for image registration.
• The derived methods offer improved accuracy and robustness, especially for groupwise tasks.
• Deep metric registration enhances performance and reduces reliance on well-registered training data.