3-D Shape Matching and Non-Rigid Correspondence for Hippocampi Based on Markov Random Fields

Area of Science:

• Medical Imaging
• Computer Vision
• Computational Anatomy

Background:

• Dense correspondence between non-rigid anatomical shapes is vital for disease diagnosis and analysis.
• Existing methods may struggle with accuracy or computational efficiency.

Purpose of the Study:

• To develop a robust framework for recovering dense correspondence between non-rigid anatomical shapes.
• To improve computational efficiency and matching accuracy in shape analysis.

Main Methods:

• A shape matching framework utilizing Markov random fields (MRFs) was proposed.
• An energy function was formulated with unary and binary terms, minimizing it via Loopy Belief Propagation (LBP).
• A sparse update technique for LBP and an expectation-maximization (EM)-like clamping approach were introduced.

Main Results:

• The sparse update for LBP demonstrated a 160x speed increase compared to standard Belief Propagation (BP).
• The EM-like clamping procedure enhanced matching accuracy, achieving a 97% matching rate on hippocampal data.
• The approach effectively overcomes the 'flip problem' and does not require pre-alignment.

Conclusions:

• The proposed MRF-based framework provides an efficient and accurate solution for non-rigid shape correspondence in anatomical objects.
• This method offers significant advantages over traditional techniques like iterative closest point (ICP) for complex anatomical shape analysis.