Related Experiment Video
Updated: Apr 25, 2026
09:33
Diffusion Tensor Magnetic Resonance Imaging in the Analysis of Neurodegenerative Diseases
Published on: July 28, 2013
27.9K
Generative Synthesis of Fractional Anisotropy Maps from T1 MRI Using Transfer Learning for White Matter Assessment in
Gyubin Kwon1, Hyunjin Kim1, Hongmin Kim1
1Department of Biomedical Engineering, Kumoh National Institute of Technology, 350-27 Gumi-Daero, Gumi, Gyeongbuk, Republic of Korea.
Brain Topography
|April 24, 2026
Summary
This study developed a transfer learning method using generative adversarial networks (GANs) to create fractional anisotropy (FA) maps from T1-weighted MRI scans. This approach offers a faster, more accessible alternative to diffusion tensor imaging (DTI) for assessing white matter integrity after stroke.
Area of Science:
- Neuroimaging
- Medical Image Analysis
- Artificial Intelligence in Medicine
Background:
- Assessing white matter integrity is crucial for predicting functional recovery post-ischemic stroke.
- Conventional MRI lacks tract-specific detail, and diffusion tensor imaging (DTI) has lengthy acquisition times.
- Developing faster, more detailed methods for white matter integrity assessment is needed for clinical stroke management.
Purpose of the Study:
- To synthesize fractional anisotropy (FA) maps from T1-weighted (T1) MRI using a generative adversarial network (GAN).
- To evaluate a transfer learning strategy (NLT+LF) against single-domain training (NLT, LT) for synthesizing FA maps.
- To determine if synthesized FA maps accurately represent white matter integrity, including lesion-affected areas.
Main Methods:
- A generative adversarial network (GAN) framework was employed to synthesize FA maps from 2.5D T1-weighted MRI inputs.
- Three models were trained: non-lesion-trained (NLT), lesion-trained (LT), and NLT fine-tuned on stroke data (NLT+LF).
- Performance was assessed using voxel-wise errors (MAE, RMSE), structural similarity (PSNR, SSIM), spatial overlap (Dice), and distributional similarity (Kullback-Leibler divergence).
Main Results:
- The NLT+LF model demonstrated significantly superior performance across all evaluated metrics compared to NLT and LT models.
- NLT+LF showed significant improvements in whole brain, white matter, and lesion regions (p < 0.001).
- The NLT+LF approach effectively preserved lesion-relevant features and anatomical fidelity, capturing degeneration patterns.
Conclusions:
- The proposed NLT+LF transfer learning framework reliably synthesizes high-fidelity FA maps from T1 MRI.
- This method offers a practical and efficient alternative to DTI for clinical assessment of white matter integrity after stroke.
- The GAN-based transfer learning approach enhances lesion-specific representation, aiding in predicting functional recovery.
