Enhanced DTCMR With Cascaded Alignment and Adaptive Diffusion.

Area of Science:

• Cardiovascular Magnetic Resonance Imaging
• Medical Image Analysis
• Computational Imaging

Background:

• Diffusion Tensor Cardiovascular Magnetic Resonance (DTCMR) is crucial for non-invasive myocardial microstructure visualization.
• Challenges in DTCMR include inconsistent breath-holds and cardiac triggering, leading to motion artifacts and inaccurate tensor fitting.
• Existing registration methods struggle with DTCMR's specific motion patterns and low signal-to-noise ratio (SNR) frames.

Purpose of the Study:

• To develop a novel deep learning framework for groupwise deformable registration in DTCMR.
• To accurately correct intra-subject inter-frame motion, including in-plane and through-plane displacements.
• To enhance the accuracy of clinical biomarker tensor estimation in DTCMR.

Main Methods:

• A novel deep learning framework incorporating tensor information for groupwise deformable registration.
• A cascaded registration branch to address in-plane and through-plane motions.
• A parallel branch for pseudo-frame generation, diffusion contrast enhancement, and template updates, guided by a refined loss function and denoising.

Main Results:

• The method achieved significantly reduced tensor fitting errors compared to traditional and deep learning methods.
• Demonstrated the lowest percentage of negative eigenvalues (0.446%) and highest R2 for HA line profiles (0.911).
• Exhibited no negative Jacobian Determinant and the shortest reference time (0.06 seconds per case).

Conclusions:

• The proposed deep learning framework effectively corrects inter-frame motion in DTCMR imaging.
• This approach significantly improves DTCMR image quality and the accuracy of tensor-derived biomarkers.
• The method shows substantial clinical potential for enhanced DTCMR diagnostics.