Training Deep Learning Based Dynamic MR Image Reconstruction Using Synthetic Fractals
Summary: arXiv:2603.29922v1 Announce Type: cross
Abstract
The purpose of this research is to investigate whether synthetically generated fractal data can be utilized to train deep learning (DL) models for dynamic MRI reconstruction. This approach aims to circumvent the privacy, licensing, and availability limitations that are commonly associated with cardiac MR training datasets.
Methods
A novel training dataset was generated using quaternion Julia fractals, which allowed for the production of 2D+time images. To simulate multi-coil MRI acquisition, paired fully sampled and radially undersampled k-space data were generated. A 3D UNet deep artifact suppression model was trained using these synthetic fractal data (referred to as F-DL) and subsequently compared with an identical model trained on traditional cardiac MRI data (denoted as CMR-DL).
Evaluation
Both models were evaluated on prospectively acquired radial real-time cardiac MRI from a cohort of 10 patients. The reconstructions were compared against two established methods: compressed sensing (CS) and low-rank deep image prior (LR-DIP). The image quality of all reconstructions was ranked, while ventricular volumes and ejection fractions were measured and compared against reference breath-hold cine MRI.
Results
- There was no significant difference in qualitative ranking between the F-DL and CMR-DL models (p=0.9).
- Both F-DL and CMR-DL outperformed the CS and LR-DIP methods in terms of image quality.
- Statistical analyses revealed comparable performance between the two deep learning models, indicating that synthetic data can effectively substitute traditional training datasets.
Conclusion
The findings of this study suggest that synthetic fractal data can be a viable alternative for training deep learning models for dynamic MRI reconstruction. The high level of performance achieved by the F-DL model, matched with that of the CMR-DL model, indicates promising potential for further applications in medical imaging where data privacy and access issues are prevalent. Future research will explore the scalability of this approach and its implications for diverse imaging scenarios.
Implications for Future Research
This research opens new avenues for the use of synthetic data in medical imaging, particularly in scenarios where acquiring real patient data is challenging. Future efforts may focus on increasing the complexity of the synthetic datasets and exploring their application in other imaging modalities beyond MRI.