Accelerating Cardiac Diffusion Tensor Imaging With a U-Net Based Model: Toward Single Breath-Hold.

Area of Science:

• Cardiovascular Imaging
• Medical Physics
• Artificial Intelligence in Medicine

Background:

• In vivo cardiac diffusion tensor imaging (cDTI) is crucial for characterizing myocardial microstructure.
• Clinical application of cDTI is hindered by technical challenges, primarily a low signal-to-noise ratio.
• Reducing scan time is essential for broader clinical adoption of cDTI.

Purpose of the Study:

• To develop a fitting-free deep learning approach for reconstructing diffusion tensors in cDTI.
• The primary goal is to reduce scan acquisition time to a single breath-hold.
• To enable faster and more efficient cardiac microstructure analysis.

Main Methods:

• A U-Net deep learning model was trained to reconstruct diffusion tensor elements from reduced cardiac diffusion tensor imaging datasets.
• Datasets were acquired with reduced breath-holds (5, 3, or 1 BH) per slice.
• Performance was compared against conventional linear least-squares (LLS) tensor fitting using metrics like fractional anisotropy (FA) and mean diffusivity (MD).

Main Results:

• The U-Net model significantly outperformed LLS tensor fitting in pixel-wise error analysis for datasets acquired in 1 or 3 breath-holds.
• Both LLS and U-Net methods showed some bias compared to reference results with reduced datasets.
• LLS showed no significant difference from reference for MD at 5 BH (P=0.38) and 3 BH (P=0.09).

Conclusions:

• Diffusion tensor prediction using a trained U-Net is a viable and promising strategy.
• This deep learning approach effectively minimizes the number of breath-holds required for clinical cDTI.
• The findings suggest a pathway towards more practical and time-efficient in vivo cardiac microstructure assessment.