Information theoretic similarity measures in non-rigid registration

Area of Science:

• Medical Imaging
• Image Registration
• Computational Anatomy

Background:

• Mutual Information (MI) and Normalised Mutual Information (NMI) are established image similarity metrics for medical image registration.
• Their application in non-rigid registration is often empirical, lacking theoretical insight into parameter influence.

Purpose of the Study:

• To derive analytical expressions for MI and NMI changes in response to local deformations.
• To implement these derived expressions as driving forces in a fluid-based non-rigid registration framework.
• To evaluate the performance of NMI-driven fluid registration using simulated multi-spectral MR brain images.

Main Methods:

• Derivation of analytical expressions for MI and NMI based on intensity histogram representations.
• Implementation of derived expressions as forces within a fluid registration algorithm.
• Empirical testing on simulated multi-spectral Magnetic Resonance (MR) brain imaging datasets.

Main Results:

• Novel expressions provide theoretical insight into Normalized Mutual Information's behavior during registration.
• The implemented fluid registration framework, driven by NMI, demonstrates functional performance.
• Performance evaluation on simulated MR brain images validates the approach.

Conclusions:

• The derived expressions offer a deeper understanding of NMI in medical image registration.
• Integrating these expressions into fluid registration frameworks enhances their operational basis.
• The study validates the use of NMI-driven fluid registration for multi-spectral MR brain images.