Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Phy-Diff: Physics-guided Hourglass Diffusion Model for Diffusion MRI Synthesis

Juanhua Zhang, Ruodan Yan, Alessandro Perelli, Xi Chen, Chao Li

TL;DR

Phy-Diff advances diffusion MRI synthesis by embedding dMRI physics into a diffusion model's forward noise evolution and by employing a query-based conditioning mechanism for q-space information, augmented with an XTRACT atlas-based tract adapter to preserve tract anatomy. The approach uses an ADC-derived noise term and $b$-vector/$b$-value cues within an HDiT backbone to generate high-fidelity, anatomically coherent dMRI across multiple $b$-values, demonstrated on the HCP S1200 dataset with strong quantitative and qualitative gains over SOTA baselines. Ablation studies confirm the contributions of physics-guided noise, q-space conditioning, and tract adaptation to performance. Overall, Phy-Diff offers a promising route toward high-quality raw dMRI synthesis with potential reductions in diffusion acquisition time and improved clinical utility.

Abstract

Diffusion MRI (dMRI) is an important neuroimaging technique with high acquisition costs. Deep learning approaches have been used to enhance dMRI and predict diffusion biomarkers through undersampled dMRI. To generate more comprehensive raw dMRI, generative adversarial network based methods are proposed to include b-values and b-vectors as conditions, but they are limited by unstable training and less desirable diversity. The emerging diffusion model (DM) promises to improve generative performance. However, it remains challenging to include essential information in conditioning DM for more relevant generation, i.e., the physical principles of dMRI and white matter tract structures. In this study, we propose a physics-guided diffusion model to generate high-quality dMRI. Our model introduces the physical principles of dMRI in the noise evolution in the diffusion process and introduce a query-based conditional mapping within the difussion model. In addition, to enhance the anatomical fine detials of the generation, we introduce the XTRACT atlas as prior of white matter tracts by adopting an adapter technique. Our experiment results show that our method outperforms other state-of-the-art methods and has the potential to advance dMRI enhancement.

Phy-Diff: Physics-guided Hourglass Diffusion Model for Diffusion MRI Synthesis

TL;DR

Phy-Diff advances diffusion MRI synthesis by embedding dMRI physics into a diffusion model's forward noise evolution and by employing a query-based conditioning mechanism for q-space information, augmented with an XTRACT atlas-based tract adapter to preserve tract anatomy. The approach uses an ADC-derived noise term and -vector/-value cues within an HDiT backbone to generate high-fidelity, anatomically coherent dMRI across multiple -values, demonstrated on the HCP S1200 dataset with strong quantitative and qualitative gains over SOTA baselines. Ablation studies confirm the contributions of physics-guided noise, q-space conditioning, and tract adaptation to performance. Overall, Phy-Diff offers a promising route toward high-quality raw dMRI synthesis with potential reductions in diffusion acquisition time and improved clinical utility.

Abstract

Diffusion MRI (dMRI) is an important neuroimaging technique with high acquisition costs. Deep learning approaches have been used to enhance dMRI and predict diffusion biomarkers through undersampled dMRI. To generate more comprehensive raw dMRI, generative adversarial network based methods are proposed to include b-values and b-vectors as conditions, but they are limited by unstable training and less desirable diversity. The emerging diffusion model (DM) promises to improve generative performance. However, it remains challenging to include essential information in conditioning DM for more relevant generation, i.e., the physical principles of dMRI and white matter tract structures. In this study, we propose a physics-guided diffusion model to generate high-quality dMRI. Our model introduces the physical principles of dMRI in the noise evolution in the diffusion process and introduce a query-based conditional mapping within the difussion model. In addition, to enhance the anatomical fine detials of the generation, we introduce the XTRACT atlas as prior of white matter tracts by adopting an adapter technique. Our experiment results show that our method outperforms other state-of-the-art methods and has the potential to advance dMRI enhancement.
Paper Structure (15 sections, 8 equations, 2 figures, 2 tables)

This paper contains 15 sections, 8 equations, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Overview.
  4. Physics-guided Noise Evolution
  5. Query-based Conditional Mapping
  6. XTRACT Condition Adaptation
  7. Experiments and Results
  8. Datasets and Preprocessing
  9. Implementation Details
  10. Comparisons with SOTA and Baselines
  11. Quantitative Results.
  12. Qualitative Results.
  13. Ablation Experiments
  14. Conclusion
  15. Disclosure of Interests.

Figures (2)

  • Figure 1: Model architecture. The physics information ADC atlas $\tilde{\mathcal{D}}_n$ interacts with the original image $b_n$ during the forward process by the function $q\left(S_{n,t}|S_{n,0},\tilde{\mathcal{D}}_n\right)$. After timesteps $T$, the image turns into the $S_{n,T}$. During the reverse process, HDiT predicts the added noise in each timestep in the pixel space, while using the b-vector, b-value as conditions via query-based conditional mapping. After training the noise prediction network, the XTRACT adapter is used to incorporate tractography structure $c_X$. By sampling from the learned pattern, the synthesized image $\hat{S_{n,0}}$ is obtained.
  • Figure 2: Examples of the generated dMRI. In each b-value block, the first row contains the ground truth and synthesized images and the second row is the error maps.