Transforming Diagnostic Imaging in Low-Resource Environments: Advanced Deep Learning Approaches for Enhancing Ultra-Low-Field MRI Quality

Introduction: In regions with limited resources, the quality of medical imaging, particularly from Ultra-Low-Field (ULF) Portable MRI scanners like the Hyperfine Swoop, is hindered by low resolution and pronounced noise. Leveraging the power of deep learning, our work aims to significantly refine ULF MRI quality, making advanced medical imaging more accessible. Our goal is to enable more accurate diagnoses of conditions, such as brain tumors, in underserved areas.

Methods: We employed three deep learning (DL) techniques to enhance ULF MRI: Upscaling: ULF MR images were used as low-resolution data, paired with high-resolution conventional MR images, which acted as the training Ground-Truth. We leveraged a pre-trained neural network with a perceptually relevant loss function derived from the early layers of a pre-trained VGG 19 model for upscaling. Second, AUTOMAP reconstructs high-quality images from undersampled k-space data, mastering the inverse Fourier transform. Lastly, FBSEM-Net integrates deep learning with simulated references for noise reduction and refined reconstruction, utilizing Mean-Square Error Loss to measure fidelity.

Results: Both ESRGAN and AUTOMAP methods exhibited rapid convergence. In contrast, the unrolled iterative process showed a more gradual convergence, beginning notably from the 6th epoch and solidifying by the 15th. Post these epochs, all three modalities stabilized, with their validation losses reaching peak levels. Notably, AUTOMAP excelled beyond the other two networks in its capability to handle intricate matrices and more effectively highlight pixel-level differences relative to the ground truth. A crucial observation was the significant influence of MR image contrast on upscaling. A combination of a greater number of frames and pronounced contrast contributed to enhanced image quality and faster outcomes, further reducing loss.

Conclusion : Our study's deep learning enhancements for ULF MRI scans show promise for equitable medical imaging in resource-limited areas, with AUTOMAP yielding notably clear details. Future work will focus on refining these techniques to further reduce computational demands and improve the accessibility of high-quality medical imaging worldwide.