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

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

2024-07-10

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.

Image and Video Processing,Computer Vision and Pattern Recognition

What problem does this paper attempt to address?

The problem that this paper attempts to solve is in diffusion magnetic resonance imaging (dMRI), how to synthesize high - quality dMRI images through generative models to reduce the high - cost dMRI data collection. Specifically, the paper proposes a physics - guided diffusion model (Phy - Diff), aiming to solve the following problems:

1. **Generate more comprehensive dMRI images**: Although the existing generative adversarial network (GAN) methods can generate dMRI images containing b - value and b - vector conditions, these methods have limitations in training stability, diversity, and detail preservation. Phy - Diff improves the generative model by introducing physical principles to generate higher - quality and more diverse dMRI images.

2. **Incorporate physical principles**: dMRI is different from other conventional MRIs, and its generation process needs to consider specific physical principles, such as the Diffusion Coefficient (DC) and the Apparent Diffusion Coefficient (ADC). Phy - Diff makes the generated dMRI images more in line with the actual physical characteristics by introducing these physical principles in the noise evolution process.

3. **Maintain the integrity of the anatomical structure**: dMRI is mainly used to study the structure of white matter fiber tracts, so the generated images need to maintain the integrity of these structures. Phy - Diff enhances the anatomical details in the white matter fiber tract area of the generated images by introducing the XTRACT atlas as prior knowledge.

4. **Multi - condition controllable generation**: In order to generate dMRI images with multiple conditions, Phy - Diff designs a query - based condition mapping mechanism, allowing flexible use of conditions such as b - vectors, b - values, and slice indices during the generation process.

In summary, the main goal of this paper is to develop a generative model that can generate high - quality, diverse, and anatomically complete dMRI images, thereby reducing the cost of dMRI data collection and improving the accuracy of clinical diagnosis.