Automating in vivo cardiac diffusion tensor postprocessing with deep learning-based segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Cardiovascular Research

Background:

• Diffusion Tensor Cardiac Magnetic Resonance (DT-CMR) imaging is crucial for assessing cardiac function.
• Manual postprocessing of DT-CMR data is time-consuming and prone to variability.
• Automated methods are needed to improve efficiency and accuracy in DT-CMR analysis.

Purpose of the Study:

• To develop and validate a fully automated postprocessing framework for in vivo DT-CMR data using deep learning.
• To enhance image registration and artifact removal in DT-CMR workflows.
• To segment the left ventricle accurately using a U-Net convolutional neural network.

Main Methods:

• A U-Net based convolutional neural network was developed and trained on 348 healthy and 144 cardiomyopathy patient DT-CMR scans.
• The U-Net was utilized to automate heart segmentation, image registration, and artifact removal.
• Data were acquired at 3 T using a STEAM-EPI sequence; processing was done in MATLAB and TensorFlow.

Main Results:

• The U-Net achieved a median Dice coefficient of 0.93 for left-ventricular myocardial segmentation.
• Image registration significantly improved with U-Net segmentation (P < .0001).
• Automated artifact identification achieved an F1 score of 0.70 compared to an expert user, with good agreement in tensor measures.

Conclusions:

• The deep learning framework successfully automated DT-CMR postprocessing, enabling real-time results and reducing manual effort.
• Automatic cardiac segmentation enhanced image registration, leading to improved diffusion tensor parameter calculations.
• This automated approach offers a promising tool for efficient and accurate cardiovascular magnetic resonance analysis.