Abstract:Deep MRI reconstruction is commonly performed with conditional models that de-alias undersampled acquisitions to recover images consistent with fully-sampled data. Since conditional models are trained with knowledge of the imaging operator, they can show poor generalization across variable operators. Unconditional models instead learn generative image priors decoupled from the operator to improve reliability against domain shifts related to the imaging operator. Recent diffusion models are particularly promising given their high sample fidelity. Nevertheless, inference with a static image prior can perform suboptimally. Here we propose the first adaptive diffusion prior for MRI reconstruction, AdaDiff, to improve performance and reliability against domain shifts. AdaDiff leverages an efficient diffusion prior trained via adversarial mapping over large reverse diffusion steps. A two-phase reconstruction is executed following training: a rapid-diffusion phase that produces an initial reconstruction with the trained prior, and an adaptation phase that further refines the result by updating the prior to minimize data-consistency loss. Demonstrations on multi-contrast brain MRI clearly indicate that AdaDiff outperforms competing conditional and unconditional methods under domain shifts, and achieves superior or on par within-domain performance.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is how to improve the performance and reliability of the model under different domain variations in accelerated magnetic resonance imaging (MRI) reconstruction. Specifically, existing conditional models rely on specific imaging operators (such as sampling patterns and coil sensitivities) during training, which leads to a decline in their performance when encountering changes in imaging operators during the testing phase. To address this issue, the authors propose an adaptive prior method based on the diffusion model - AdaDiff, aiming to improve the reconstruction performance in cross - domain situations by adaptively adjusting the diffusion prior. ### Background and Motivation - **Limitations of Existing Methods**: - **Conditional Models**: These models need to know the imaging operator during training, so when the imaging operator changes (for example, changes in acceleration rate, sampling density, or the number of coils), their performance will decline significantly. - **Unconditional Models**: These models avoid dependence on the imaging operator by generating image priors, but the static diffusion prior still performs poorly when dealing with cross - domain changes. ### Proposed Method - **AdaDiff**: This method combines a fast diffusion process and adaptive prior adjustment to improve the reconstruction performance under different domain variations. - **Fast Diffusion**: A large - step - size diffusion process is achieved through an adversarial mapper, thereby accelerating the image generation speed. - **Adaptive Prior Adjustment**: During the inference stage, the diffusion prior is updated to minimize the data consistency loss, so as to better adapt to the distribution of the test data. ### Main Contributions 1. **First Proposed Adaptive Prior Method Based on Diffusion Model**: AdaDiff is the first method to use an adaptive diffusion prior in accelerated MRI reconstruction. 2. **Efficient Fast Diffusion Process**: A large - step - size diffusion process is achieved through an adversarial mapper, which significantly improves the speed of image generation. 3. **Prior Adaptation in the Inference Stage**: By updating the diffusion prior in the inference stage, the performance and reliability of the model under different domain changes are improved. ### Method Details - **Fast Diffusion Process**: - Use a diffusion process with a large step size \( k \), and the formulas are as follows: \[ x_t = \sqrt{1 - \gamma_t} x_{t - k} + \sqrt{\gamma_t} z \] \[ q(x_t|x_{t - k})=\mathcal{N}\left(x_t; \sqrt{1 - \gamma_t} x_{t - k}, \gamma_t I\right) \] - Since the large step size breaks the normal assumption, an adversarial mapper is introduced to implicitly capture the conditional distribution of the inverse diffusion step. - **Adaptive Prior Adjustment**: - Initial Reconstruction Stage: Quickly generate an initial reconstructed image by alternately performing data consistency projection and inverse diffusion projection. - Prior Adjustment Stage: Minimize the data consistency loss by updating the generator parameter \( \theta_G \), and the formula is as follows: \[ \theta^*_G=\arg\min_{\theta_G}\|A G_{\theta_G}(x_{\text{init}}, 0)-y\|_1 \] ### Experimental Results - **Multi - contrast Brain MRI Reconstruction**: Experiments were carried out on the IXI and fastMRI datasets to verify the superior performance of AdaDiff in cross - domain situations. - **Comparison with Existing Methods**: AdaDiff significantly outperforms existing conditional and unconditional models in cross - domain situations, and also achieves comparable or better performance in the same - domain situation. ### Conclusion AdaDiff effectively addresses the challenges brought by cross - domain changes in accelerated MRI reconstruction by combining the fast diffusion process and adaptive prior adjustment, and improves the robustness and performance of the model.

Adaptive Diffusion Priors for Accelerated MRI Reconstruction