Unifying encoding of spatial information in mutual information for nonrigid registration

Area of Science:

• Medical image analysis
• Computer vision
• Computational anatomy

Background:

• Conventional mutual information (MI) struggles with nonrigid registration in the presence of bias fields.
• Encoding spatial information into MI is a promising approach to overcome these limitations.

Purpose of the Study:

• To propose a unified method for encoding spatial information into the joint probability distribution function (PDF) for improved nonrigid registration.
• To demonstrate the unification of regional MI, conditional MI, and conventional MI PDF computation.

Main Methods:

• The joint probability distribution function (PDF) is computed using local volumes while incorporating global intensity information.
• Two categories of registration methods were derived based on this unified PDF computation.
• These methods were applied to various medical image registration tasks.

Main Results:

• The proposed computation method successfully unifies PDF computation for different MI variants.
• Experimental results show significant improvements in registration accuracy across both derived method categories.
• The approach effectively enhances nonrigid registration performance, especially in challenging scenarios like bias fields.

Conclusions:

• The unified PDF computation method effectively incorporates spatial information for robust nonrigid registration.
• This approach offers a significant advancement in medical image registration, particularly for overcoming bias field challenges.
• The derived methods demonstrate broad applicability and improved performance in diverse registration tasks.