NAF-GAN: Anatomically Constrained GAN for Ultra-Low-Field MRI Enhancement

High-quality diagnostic imaging in resource-limited settings remains a significant challenge due to the limited accessibility of high-field Magnetic Resonance Imaging (MRI). The Ultra-Low-Field MRI Image Enhancement Challenge (ULF-EnC) addresses this issue by leveraging paired 64mT and 3T MRI scans to develop algorithms that enhance low-field image quality to high-field standards. In this paper, we propose an end-to-end NAF-GAN framework, a generative adversarial network based on the Nonlinear Activation Free Network (NAFNet) architecture, specifically designed for this task. Our model learns the complex translation from 64mT to 3T MRI sequences, improving image clarity and contrast while preserving the anatomical fidelity essential for clinical diagnosis. Both the patch-level adversarial loss and the segmentation-based Dice loss are designed to capture high-frequency details and preserve anatomical structures. The proposed approach achieves strong performance across T1, T2, and FLAIR contrasts, underscoring its potential to enable cost-effective and accessible MRI solutions for mobile and global healthcare applications.