Generative Synthesis of Fractional Anisotropy Maps from T1 MRI Using Transfer Learning for White Matter Assessment in...

Assessment of white matter integrity is critical for predicting functional recovery after ischemic stroke. However, conventional magnetic resonance imaging (MRI) cannot capture tract-specific microstructural changes, and diffusion tensor imaging (DTI) is limited by prolonged acquisition times. This study aimed to synthesize fractional anisotropy (FA) maps from routinely acquired T1-weighted (T1) MRI using 2.5D inputs within a generative adversarial network (GAN) framework. Specifically, our primary objective was to evaluate the relative efficacy of a proposed transfer learning strategy compared to single-domain training approaches. T1-FA paired data from 375 cognitively normal participants (832 images) from the Alzheimer's Disease Neuroimaging Initiative served as the non-lesion dataset, while longitudinal MRI data from 69 ischemic stroke patients (236 images) were from a single-center cohort. Three models were evaluated: the non-lesion-trained (NLT) model trained on non-lesion data, the lesion-trained (LT) model trained on stroke data, and the NLT model further fine-tuned on the stroke dataset (NLT + LF). Model performance was evaluated using voxel-wise errors (mean absolute error (MAE) and root mean square error (RMSE)), structural similarity (peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM)), spatial overlap (Dice coefficient (Dice)), and distributional similarity (Kullback-Leibler divergence). Bonferroni-corrected paired t-tests showed that NLT + LF showed significantly better performance than NLT across all evaluated regions, including the whole brain, white matter, and lesions (all p < 0.001). Compared with LT, NLT + LF showed superior performance for all metrics at p < 0.001, except for lesion-region Dice and RMSE, which remained significant at p < 0.01. The preservation of lesion-relevant features and anatomical fidelity, together with the capture of degeneration patterns, accompanied these gains. Overall, the proposed NLT + LF approach improved lesion-specific representation and established that high-fidelity FA maps can be reliably synthesized from T1 MRI. This transfer learning framework offers a practical alternative to DTI for clinical stroke assessment.