CNN-Based Cardiac Motion Extraction to Generate Deformable Geometric Left Ventricle Myocardial Models from Cine MRI.

Area of Science:

• Medical Imaging
• Computational Biology
• Machine Learning

Background:

• Patient-specific left ventricle (LV) myocardial models are crucial for clinical applications in diagnosis and treatment planning.
• Cine cardiac magnetic resonance (MR) imaging offers high-resolution data for reconstructing LV geometric models.
• Deep learning advancements enable accurate cardiac chamber segmentation and unsupervised image registration for motion estimation.

Purpose of the Study:

• To develop a deep learning-based framework for creating patient-specific LV myocardial models from cine cardiac MR images.
• To utilize deep learning for accurate segmentation and motion estimation in cardiac MR datasets.
• To generate dynamic LV myocardial volume meshes throughout the cardiac cycle.

Main Methods:

• A deep learning framework using a VoxelMorph-based convolutional neural network (CNN) was proposed.
• The CNN estimated deformation fields to propagate end-diastole (ED) frame meshes to other cardiac phases.
• Models were compared against manually segmented and traditionally registered models, including log barrier-based mesh warping (LBWARP).

Main Results:

• The deep learning framework successfully generated patient-specific LV myocardial models.
• CNN-propagated models showed comparable or improved accuracy against traditional methods.
• Dynamic LV myocardial volume meshes were generated and validated across the cardiac cycle.

Conclusions:

• The proposed deep learning framework provides an effective method for generating patient-specific LV myocardial models.
• This approach enhances the potential for improved clinical diagnosis and personalized treatment strategies.
• The framework demonstrates the capability of deep learning in cardiac image analysis and modeling.