Improving robustness of automatic cardiac function quantification from cine magnetic resonance imaging using...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Cardiovascular Research

Background:

• Cardiac function quantification from MRI is crucial but lacks full automation in clinical practice.
• Limited training data, especially for cardiovascular disease phenotypes, hinders model development.
• Existing methods struggle with imbalanced datasets, predominantly featuring healthy subjects.

Purpose of the Study:

• To synthetically generate short-axis CINE MRI data using a generative adversarial model.
• To expand existing datasets with more cases of reduced ejection fraction.
• To develop a deep learning model for predicting cardiac function without explicit segmentation.

Main Methods:

• Generative adversarial networks (GANs) were employed to synthesize short-axis CINE MRI data.
• A convolutional neural network (CNN) was developed to predict left ventricle end-diastolic and end-systolic volumes.
• The CNN model was pre-trained on synthetically generated data to address dataset imbalance.

Main Results:

• The CNN model demonstrated superior accuracy in predicting left ventricle volumes compared to state-of-the-art segmentation methods.
• Pre-training with synthetic data significantly improved prediction accuracy on imbalanced datasets.
• The approach enables implicit ejection fraction calculation without explicit segmentation.

Conclusions:

• Synthetic MRI data generation can effectively augment limited clinical datasets.
• Deep learning models pre-trained on synthetic data enhance cardiac function quantification accuracy.
• This method offers a promising pathway toward automated cardiac MRI analysis for diverse cardiovascular conditions.