Diffusion Bridge Models for 3D Medical Image Translation

TL;DR

This work introduces diffusion bridge models to translate 3D brain images between T1-weighted MRI and diffusion tensor imaging FA maps, addressing the practical burden of DTI acquisition. The approach explicitly models the conditional distribution $\pi_{0|1}(x_0|x_1)$ through a bridge process, using a 3D ADM UNet within a diffusion framework and training with a denoising objective. Extensive experiments on ADNI-derived data demonstrate high perceptual fidelity (MS-SSIM), strong pixel-level agreement (PSNR), and distributional similarity (MMD) between real and synthetic DTI FA images, with downstream tasks (sex and AD classification) showing competitive performance when using synthetic data. The method offers a promising pathway for data augmentation and cross-modality neuroimaging analyses, with potential clinical impact, while acknowledging limitations related to dataset size and diversity and suggesting avenues for extending to full diffusion tensors and higher-resolution translations.

Abstract

Diffusion tensor imaging (DTI) provides crucial insights into the microstructure of the human brain, but it can be time-consuming to acquire compared to more readily available T1-weighted (T1w) magnetic resonance imaging (MRI). To address this challenge, we propose a diffusion bridge model for 3D brain image translation between T1w MRI and DTI modalities. Our model learns to generate high-quality DTI fractional anisotropy (FA) images from T1w images and vice versa, enabling cross-modality data augmentation and reducing the need for extensive DTI acquisition. We evaluate our approach using perceptual similarity, pixel-level agreement, and distributional consistency metrics, demonstrating strong performance in capturing anatomical structures and preserving information on white matter integrity. The practical utility of the synthetic data is validated through sex classification and Alzheimer's disease classification tasks, where the generated images achieve comparable performance to real data. Our diffusion bridge model offers a promising solution for improving neuroimaging datasets and supporting clinical decision-making, with the potential to significantly impact neuroimaging research and clinical practice.

Figures (4)

• Figure 1: Overall framework of diffusion bridge models for 3D medical image translation.
• Figure 2: Image translation from T1 to FA with 3 subjects. The table presents four types of images for three different subjects: true T1 images, synthetic FA images, and true FA images. To illustrate the approach, we include example images from participants representing three groups: healthy elderly controls, individuals with mild cognitive impairment (MCI), and those with Alzheimer's disease (dementia).
• Figure 3: 2D MS-SSIM between real and synthetic FA images across different views. $\mu$ represents the mean of MS-SSIM across all slices. The MS-SSIM value for each slice is computed as the mean across 167 subjects in the test dataset. The synthetic FA images are generated by the same pretrained diffusion bridge model with different samplers: ODE sampler ($\eta = 0$), SDE sampler ($\eta = 1.0$).
• Figure 4: 3D MS-SSIM, PSNR and MMD evaluation between real and synthetic FA images across $167$ subjects. The synthetic FA images are generated by the same pretrained diffusion bridge model with different samplers: ODE sampler ($\eta = 0$), SDE sampler ($\eta = 1.0$).