Cross-Modality Whole-Heart MRI Reconstruction with Deep Motion Correction and Super-Resolution.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Research

Background:

• Cardiac MRI suffers from motion artifacts and misalignment, leading to inaccurate 3D reconstructions and functional assessments.
• High-resolution MRI requires long scan times, increasing patient burden and potential risks.

Purpose of the Study:

• To develop a deep learning framework (DeepWHR) for motion correction and super-resolution whole-heart reconstruction from cardiac MRI.
• To improve the anatomical accuracy and resolution of cardiac structures derived from MRI data.

Main Methods:

• DeepWHR learns cardiac structure priors from CT data to reconstruct MRI data with motion correction and super-resolution.
• A deep motion correction model trained on CT anatomy data ensures structural coherence.
• An implicit neural representation module enables multi-scale super-resolution reconstruction.

Main Results:

• DeepWHR successfully restores spatial coherence and anatomical consistency to cardiac MRI data.
• The framework generates high-fidelity label representations suitable for downstream cardiac applications.
• Experiments on the CARE2024 WHS dataset validate the method's effectiveness.

Conclusions:

• DeepWHR transforms sparse, misaligned 2D MRI data into anatomically coherent, high-resolution 3D cardiac models.
• This enhancement improves the reliability of cardiac models for clinical applications.
• The framework addresses key limitations in current cardiac MRI acquisition and reconstruction.