3T to 7T-like MRI Translation Using Attention-Guided Multi-Scale CycleGAN

This paper introduces the Attention-Guided Multi-Scale CycleGAN (AGMS-CycleGAN) for generating 7T-like MRI images from 3T scans, utilizing an attention mechanism to focus on critical regions. The framework integrates transformer modules with hybrid losses—combining texture preservation and structural coherence—to enhance anatomical fidelity. Evaluated on the UNC 3T-7T benchmark using multi-planar data (axial (A<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">x</inf>), coronal (C<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">r</inf>), sagittal (S<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">g</inf>)) from 10 subjects, two training strategies were employed: the "Merged Planes", where the planes were combined (A<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">x</inf>C<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">r</inf>S<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">g</inf>), and the "Separate Planes" approach, where each plane was processed individually. AGMS-CycleGAN outperforms CycleGAN, MSR-CycleGAN, and SRGAN, achieving a PSNR of 28.86 and an SSIM of 0.87, with the best results observed in the sagittal plane. This work underscores the efficacy of attention mechanisms in bridging the resolution gap between 3T and 7T MRI while maintaining compatibility with unpaired training paradigms.