PT-CycleGAN: Progressive Transformer Based CycleGAN for 3 T to 7 T Like MRI Synthesis.
Authors
Affiliations (2)
Affiliations (2)
- Department of Electronics and Communication Engineering, Indian Institute of Information Technology (IIIT) Guwahati, Guwahati, Assam, India (F.B., S.B.). Electronic address: [email protected].
- Department of Electronics and Communication Engineering, Indian Institute of Information Technology (IIIT) Guwahati, Guwahati, Assam, India (F.B., S.B.).
Abstract
Existing Generative Adversarial Networks (GANs) for 3 T to 7 T MRI synthesis often suffer from structural instability and an inability to preserve fine grained anatomical textures due to rigid architectures. This study proposes PT-CycleGAN, a progressive transformer based framework designed to capture the hierarchical nature of neuroanatomy while maintaining training stability. The proposed model utilizes a bidirectional CycleGAN architecture with a novel generator that dynamically scales attention complexity by increasing the number of attention heads across layers. A staged training strategy was implemented to stabilize global geometry before synthesizing high frequency details. Validation was conducted on T1 weighted and T2 weighted scans from two benchmark datasets (UNC and Hippocampal) using plane wise (A<sub>x</sub>, C<sub>r</sub>, S<sub>g</sub>) processing. Performance was evaluated using PSNR, SSIM, and SHAP based interpretability analysis. PT-CycleGAN significantly outperformed state of the art baselines, including SRGAN and MSR-CycleGAN. The model achieved a peak PSNR of 34.67 dB and an SSIM of 0.95. Qualitative assessment confirmed the recovery of sharp gray white matter interfaces and submillimeter cortical details. SHAP analysis verified that the synthesis process was guided by clinically salient features, specifically cerebrospinal fluid and gray matter regions. By implementing dynamic attention scaling and a synchronized staged training schedule, PT-CycleGAN provides a robust solution for high fidelity MRI translation. This approach effectively addresses architectural rigidity in traditional models, establishing a more reliable framework for ultra high field neuroimaging synthesis.