Physics-Coupled Synthetic Data Generation Method Using Natural Images Enabling In-vivo Data-Free Complex MRI Denoising

Motivation: While deep learning has advanced MRI denoising, it still relies heavily on in-vivo data, which presents challenges due to the lack of strictly noise-free ground truth and the difficulty of obtaining large, diverse in-vivo datasets. Goal(s): We aim to synthesize complex MRI-like data from natural images to enable robust, high-quality denoising with reduced dependence on in-vivo data. Approach: Our physics-coupled method synthesizes MRI-like data from natural images by embedding structural information into the phase, producing realistic training data without in-vivo samples. Results: This approach achieves denoising performance comparable to in-vivo-trained models across diverse conditions, enabling effective neural network-based denoising without extensive in-vivo datasets. Impact: Our method generates physics-coupled synthetic data from natural images, enabling effective complex MRI denoising without in-vivo data, achieving performance on par with in-vivo-trained models. This approach reduces dependence on large in-vivo datasets and addresses practical challenges in in-vivo data collection.