Automatic myocardium strain quantification in MR synthetic images with Deep Leaning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Accurate myocardium strain quantification is crucial for diagnosing and monitoring cardiac diseases.
• Current methods using brightness pattern tracking in cine-MR images are limited by myocardium homogeneity, reducing motion and strain estimation accuracy.

Purpose of the Study:

• To investigate the feasibility of quantifying myocardium strain in cardiac resonance synthetic images using motion estimated by a neural network.
• To evaluate the performance of a convolutional neural network for motion estimation in cardiac MRI.

Main Methods:

• Utilized a neural network, Pyramid, Warping, and Cost Volume (PWC), to generate motion data from synthetic cardiac resonance images.
• Quantified myocardium strain in both radial and circumferential directions based on the neural network-generated motion.
• Compared the neural network's strain estimation accuracy against two classical motion tracking methods.

Main Results:

• The neural network-based method achieved a mean average error of 12.30% ± 6.50% for radial strain.
• Circumferential strain estimation yielded a mean average error of 1.20% ± 0.61%.
• The proposed method demonstrated superior performance compared to two evaluated classical methods.

Conclusions:

• This work demonstrates the feasibility of estimating myocardium strain using motion estimated by a convolutional neural network.
• Neural network-based motion estimation offers a promising advancement for accurate cardiac strain quantification.
• The findings suggest potential for improved diagnosis and monitoring of cardiac diseases.