Ultra-Low Field Magnetic Resonance Image Enhancement based on Deep-Learning Method

In recent years, there has been growing interest in ultra-low-field (ULF) magnetic resonance imaging (MRI) technology aimed at addressing the limitations of high-field MRI systems, including high cost, immobility, and maintenance challenges. Under ultra-low background magnetic field conditions (<100 mT), MRI signals are weaker, and the overall signal-to-noise ratio of the k-space data used for imaging is poor. Additionally, the image resolution is often lower to shorten the scan time. These two factors significantly impact image quality. To address these issues, it is essential to enhance image quality during the image reconstruction and post-processing phases. This paper initially processes the k-space data into a low-rank matrix and performs preliminary noise reduction using the Singular Value Decomposition (SVD) algorithm. After image reconstruction, a deep convolutional denoising network is applied for denoising in the image domain. Finally, the denoised image data undergo super-resolution using bicubic interpolation, resulting in enhanced ULF MRI images. These methods, trained on 1.5T high-field MRI image data, are applied to 54 mT ULF MRI brain imaging. The processed image quality is significantly superior to the original images. This advancement is expected to enhance the value and significance of ULF MRI in clinical applications and promote the expansion of MRI technology.