Revolutionizing Brain Tumor Imaging: Generating Synthetic 3D FA Maps from T1-Weighted MRI using CycleGAN Models

TL;DR

This work targets the misalignment between FA maps and tractography atlases by synthesizing 3D FA and DEC maps directly from T1-weighted MRI using a 3D CycleGAN trained on unpaired data. The model integrates an adversarial loss, cycle-consistency loss, and a Cor-Coe loss to preserve structure, with a joint objective L that balances realism and content fidelity, evaluated via $SSIM$, $MS$-$SSIM$, and $PSNR$, along with radiologist assessments. Results show high-fidelity synthesis in healthy brain regions and robust performance in tumour areas, with transfer learning from healthy to tumour data improving tumour-map quality and enabling cross-modal applicability including downstream segmentation. The findings suggest AI-driven diffusion-metric synthesis from standard MRI can streamline workflows, reduce additional scan requirements, and support downstream neuroimaging analyses, though refinements are needed to better preserve local microstructural details and anatomical precision.

Abstract

Fractional anisotropy (FA) and directionally encoded colour (DEC) maps are essential for evaluating white matter integrity and structural connectivity in neuroimaging. However, the spatial misalignment between FA maps and tractography atlases hinders their effective integration into predictive models. To address this issue, we propose a CycleGAN based approach for generating FA maps directly from T1-weighted MRI scans, representing the first application of this technique to both healthy and tumour-affected tissues. Our model, trained on unpaired data, produces high fidelity maps, which have been rigorously evaluated using Structural Similarity Index (SSIM) and Peak Signal-to-Noise Ratio (PSNR), demonstrating particularly robust performance in tumour regions. Radiological assessments further underscore the model's potential to enhance clinical workflows by providing an AI-driven alternative that reduces the necessity for additional scans.

Figures (6)

• Figure 1: The schematic of the 3D CycleGAN architecture.
• Figure 2: Bland-Altman plot comparing the generated FA maps with the ground truth for both healthy and tumor cases (encompassing results of learning from scratch and transfer learning approaches).
• Figure 3: Quantitative evaluation of the generated FA maps using PSNR. Higher PSNR values indicate lower reconstruction error and better image quality.
• Figure 4: Comparison of generated images with and without tumour masking based on PSNR evaluation. The left subfigure provides an illustration of the tumour masking conditions. The right subfigure presents the PSNR distribution for cases where tumor masking is applied or not, under both transfer learning and direct learning settings. Additionally, the bar plot in the top right corner visualizes the distribution of case-wise PSNR differences between transfer learning and direct learning (transfer – direct), providing insights into the impact of transfer learning on reconstruction quality.
• Figure 5: Comparison of generated image quality between the tumour region, the region after tumour masking, and the whole brain image. The evaluation tracks the differences in image quality across these distinct areas.