Deep-learning-based image reconstruction with limited data: generating synthetic raw data using deep learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning excels at reconstructing accelerated MRI scans using extensive raw data.
• Limited availability of raw MRI data poses a challenge for training deep learning models.
• Synthetic data generation offers a potential solution to augment small datasets.

Purpose of the Study:

• To investigate the efficacy of synthetic data generation for improving accelerated MRI reconstruction.
• To assess the impact of synthetic data on deep learning model performance with varying dataset sizes.
• To explore the use of magnitude-only datasets for synthetic raw data creation.

Main Methods:

• An adversarial auto-encoder was employed to generate synthetic phase and coil sensitivity maps from magnitude images.
• Synthetic raw MRI data was created by combining generated maps with magnitude images.
• Deep learning reconstruction networks were trained using varying quantities of real and synthetic data (20-160 scans) for a fourfold accelerated MR task.

Main Results:

• Training with synthetic data reduced reconstruction errors, particularly with smaller training sets (up to 7.5% decrease in Mean Absolute Error).
• For larger training sets, synthetic data led to a slight increase in MAE (up to 2.6%).
• Performance gains were observed when synthetic data was used to supplement limited real raw data.

Conclusions:

• Synthetic raw data generation effectively improves MRI reconstruction quality in data-limited scenarios.
• This approach enables the utilization of more accessible magnitude-only datasets, overcoming the scarcity of raw MRI data.
• Synthetic data offers a valuable strategy for enhancing deep learning-based accelerated MRI reconstruction.