Training Deep Learning Based Dynamic MR Image Reconstruction Using Synthetic Fractals

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Data Science

Background:

• Dynamic MRI reconstruction faces challenges due to privacy, licensing, and data availability issues with clinical datasets.
• Deep learning (DL) models show promise for improving MRI reconstruction quality and efficiency.
• Synthetic data generation offers a potential solution to overcome limitations of real-world clinical data.

Purpose of the Study:

• To evaluate the efficacy of synthetically generated fractal data in training deep learning models for dynamic MRI reconstruction.
• To assess if fractal-based DL models can achieve comparable performance to models trained on natural videos or actual cardiac MRI data.
• To explore the potential of synthetic data to address privacy, licensing, and availability constraints in cardiac MRI training datasets.

Main Methods:

• A synthetic training dataset was created using quaternion Julia fractals, simulating 2D+time MRI acquisition with undersampled k-space data.
• A 3D UNet deep artifact suppression model was trained using fractal data (F-DL), natural videos (NV-DL), and cardiac MRI data (CMR-DL).
• Model performance was evaluated on prospectively acquired real-time cardiac MRI, comparing reconstructions against compressed sensing (CS) and low-rank deep image prior (LR-DIP), and assessing image quality and clinical measurements.

Main Results:

• No significant difference in qualitative image quality ranking was observed between F-DL, NV-DL, and CMR-DL models (p > 0.75).
• All DL models significantly outperformed CS and LR-DIP methods (p < 0.05) in reconstruction quality.
• Ventricular volumes and ejection fraction derived from F-DL reconstructions were comparable to CMR-DL, with no significant bias and acceptable limits of agreement against reference cine MRI.

Conclusions:

• Deep learning models trained on synthetic fractal data can reconstruct real-time cardiac MRI with image quality and clinical utility comparable to models trained on actual cardiac MRI data.
• Fractal training data represent an open, scalable, and privacy-compliant alternative to clinical datasets for dynamic MRI reconstruction.
• This approach may facilitate the development of more generalizable deep learning models for dynamic MRI applications.