Related Experiment Video
Updated: Aug 5, 2026
08:51
Magnetic Resonance Imaging of Multiple Sclerosis at 7.0 Tesla
Published on: February 19, 2021
PT-CycleGAN: Progressive Transformer Based CycleGAN for 3 T to 7 T Like MRI Synthesis
Franklin Burhagohain1, Shovan Barma1
1Department of Electronics and Communication Engineering, Indian Institute of Information Technology (IIIT) Guwahati, Guwahati, Assam, India (F.B., S.B.).
Academic Radiology
|July 27, 2026
Summary
PT-CycleGAN, a novel progressive transformer framework, enhances Magnetic Resonance Imaging (MRI) synthesis by improving structural stability and preserving fine anatomical textures. This advanced Generative Adversarial Network (GAN) model offers superior performance for ultra-high field neuroimaging.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Neuroscience
Background:
- Existing Generative Adversarial Networks (GANs) for 3T to 7T MRI synthesis exhibit structural instability and texture degradation due to rigid architectures.
- Capturing hierarchical neuroanatomy and maintaining training stability are critical challenges in high-field MRI synthesis.
Purpose of the Study:
- To introduce PT-CycleGAN, a progressive transformer-based framework for stable and high-fidelity MRI synthesis.
- To address the limitations of existing GANs in preserving anatomical details and ensuring structural integrity.
Main Methods:
- Utilized a bidirectional CycleGAN architecture with a novel generator featuring dynamic attention scaling.
- Implemented a staged training strategy to stabilize global geometry before high-frequency detail synthesis.
- Validated on T1/T2 weighted scans from benchmark datasets (UNC, Hippocampal) using plane-wise processing and evaluated with PSNR, SSIM, and SHAP analysis.
Main Results:
- PT-CycleGAN significantly outperformed state-of-the-art baselines (SRGAN, MSR-CycleGAN).
- Achieved a peak PSNR of 34.67 dB and an SSIM of 0.95.
- Qualitative assessment confirmed recovery of sharp gray/white matter interfaces and submillimeter cortical details; SHAP analysis indicated synthesis guided by clinically relevant features.
Conclusions:
- PT-CycleGAN offers a robust solution for high-fidelity MRI translation through dynamic attention scaling and staged training.
- Effectively overcomes architectural rigidity in traditional models for ultra-high field neuroimaging synthesis.
- Establishes a more reliable framework for advanced MRI synthesis applications.
