Diffusion Models To Predict 3D Late Mechanical Activation From Sparse 2D Cardiac MRIs.

Area of Science:

• Medical imaging
• Computational cardiology
• Artificial intelligence

Background:

• Accurate 3D late mechanical activation (LMA) maps of the left ventricle are vital for cardiac resynchronization therapy (CRT) in heart failure.
• Current deep learning models for LMA reconstruction from 2D cardiac MRI often neglect myocardial shape, limiting accuracy.

Purpose of the Study:

• To develop a novel shape-constrained diffusion model for improved 3D LMA map reconstruction from sparse 2D cardiac MRI.
• To leverage object shape priors to guide the reconstruction process, enhancing accuracy over intensity-based methods.

Main Methods:

• A joint learning network was developed to learn a mean myocardial shape under deformation models.
• Reconstructed images are treated as deformed variants of the learned mean shape.
• The model utilizes shape priors alongside image intensity for 3D reconstruction.

Main Results:

• The proposed shape-constrained diffusion model demonstrated superior performance in reconstructing 3D LMA maps.
• Experimental results showed improved accuracy compared to state-of-the-art deep learning reconstruction models.
• Validation was performed on a public 3D myocardium mesh dataset.

Conclusions:

• Shape-constrained diffusion models offer a more accurate approach to reconstructing 3D LMA maps from limited 2D cardiac MRI data.
• This method has the potential to improve the prediction of late activating regions and optimize CRT site selection.
• Integrating shape priors enhances the robustness and accuracy of deep learning-based cardiac image reconstruction.