A Convolutional Neural Network-based Deformable Image Registration Method for Cardiac Motion Estimation from Cine...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Research

Background:

• Deformable registration of 3D cine cardiac magnetic resonance (CMR) images is crucial for analyzing cardiac motion.
• Slice misalignment in CMR images can hinder accurate registration and subsequent analysis.
• Accurate segmentation of cardiac structures like the left ventricle (LV) and right ventricle (RV) is essential for quantitative assessment.

Purpose of the Study:

• To develop an unsupervised deep learning framework for deformable registration of 3D cine CMR images.
• To improve the accuracy of cardiac structure segmentation and motion estimation.
• To introduce a pipeline for patient tractography estimation using CNN-based cardiac motion estimation.

Main Methods:

• Implemented an unsupervised deep learning framework with a Laplacian-based operator for smoothing loss.
• Utilized a U-Net model for segmenting LV blood-pool, LV myocardium, and RV blood-pool to correct slice misalignment.
• Applied the registration deformation field to warp segmentation labels across cardiac phases using the Automated Cardiac Diagnosis Challenge (ACDC) dataset.

Main Results:

• Achieved high mean Dice scores for segmented structures: 94.84% (LV blood-pool), 85.22% (LV myocardium), and 84.36% (RV blood-pool).
• Obtained low mean Hausdorff distances: 2.74 mm (LV blood-pool), 5.88 mm (LV myocardium), and 9.04 mm (RV blood-pool).
• Demonstrated a novel pipeline for estimating patient tractography based on the developed CNN-based cardiac motion estimation.

Conclusions:

• The proposed unsupervised deep learning framework effectively performs deformable registration of 3D cine CMR images.
• The method significantly improves the accuracy of segmenting and tracking cardiac structures throughout the cardiac cycle.
• The framework provides a robust foundation for advanced cardiac analysis, including motion estimation and tractography.