Using synthetic data generation to train a cardiac motion tag tracking neural network

Area of Science:

• Medical Imaging
• Cardiovascular Imaging
• Machine Learning in Medicine

Background:

• Cardiac MRI tag tracking is crucial for assessing heart function.
• Manual tracking is time-consuming and prone to inter-observer variability.
• Developing automated, accurate methods is essential for clinical translation.

Purpose of the Study:

• To develop and validate a CNN-based automated method for cardiac MRI tag tracking.
• To assess the accuracy and precision of the proposed method using synthetic and in vivo data.

Main Methods:

• A synthetic data simulator was created using natural images, Bloch equation simulation, varied tissue properties, and programmed motion.
• A Convolutional Neural Network (CNN) was trained on the generated synthetic data.
• Validation was performed using an analytical deforming cardiac phantom and in vivo human cardiac MRI data.

Main Results:

• The CNN method demonstrated accurate results in the analytical phantom for signal-to-noise ratios (SNR) > 10, with displacement error < 0.3 mm.
• Excellent agreement was observed in vivo for tag locations (mean displacement difference = -0.02 pixels) and calculated cardiac circumferential strain (mean difference = 0.006).

Conclusions:

• Automated cardiac MRI tag tracking using a CNN trained on synthetic data is both accurate and precise.
• This method holds promise for efficient and reliable assessment of cardiac mechanics.