Deformable Medical Image Registration With Effective Anatomical Structure Representation and Divide-and-Conquer...

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Deformable medical image registration (DMIR) performance is enhanced by effective Region of Interest (ROI) representation and independent alignment.
• Current learning-based DMIR methods have limitations: unsupervised methods ignore ROI representation, and weakly-supervised methods rely heavily on label constraints.

Purpose of the Study:

• To introduce a novel weakly-supervised ROI-based registration approach, EASR-DCN, that achieves independent ROI alignment without requiring labels.
• To represent medical images using effective ROIs and align them independently to overcome limitations of existing DMIR methods.

Main Methods:

• Utilized a Gaussian mixture model for intensity analysis to represent images via multiple ROIs with distinct intensities.
• Proposed a novel Divide-and-Conquer Network (DCN) to process ROIs through separate channels for independent feature alignment.
• Integrated sub-deformation fields to generate a comprehensive displacement vector field.

Main Results:

• EASR-DCN demonstrated superior accuracy and deformation reduction efficacy across three MRI and one CT datasets.
• Achieved significant Dice score improvements compared to VoxelMorph: 10.31% (brain MRI), 13.01% (cardiac MRI), and 5.75% (hippocampus MRI).

Conclusions:

• EASR-DCN offers a promising approach for accurate and efficient deformable medical image registration.
• The method's ability to perform independent ROI alignment without labels highlights its potential for clinical applications.