Abstract:Deep learning-based 3D imaging, in particular magnetic resonance imaging (MRI), is challenging because of limited availability of 3D training data. Therefore, 2D diffusion models trained on 2D slices are starting to be leveraged for 3D MRI reconstruction. However, as we show in this paper, existing methods pertain to a fixed voxel size, and performance degrades when the voxel size is varied, as it is often the case in clinical practice. In this paper, we propose and study several approaches for resolution-robust 3D MRI reconstruction with 2D diffusion priors. As a result of this investigation, we obtain a simple resolution-robust variational 3D reconstruction approach based on diffusion-guided regularization of randomly sampled 2D slices. This method provides competitive reconstruction quality compared to posterior sampling baselines. Towards resolving the sensitivity to resolution-shifts, we investigate state-of-the-art model-based approaches including Gaussian splatting, neural representations, and infinite-dimensional diffusion models, as well as a simple data-centric approach of training the diffusion model on several resolutions. Our experiments demonstrate that the model-based approaches fail to close the performance gap in 3D MRI. In contrast, the data-centric approach of training the diffusion model on various resolutions effectively provides a resolution-robust method without compromising accuracy.

What problem does this paper attempt to address?

### What problem does this paper attempt to solve? This paper aims to solve the problem of resolution robustness in **3D magnetic resonance imaging (MRI) reconstruction**. Specifically, existing 3D MRI reconstruction methods based on diffusion models usually assume a fixed voxel size during training. When the voxel size changes in practical applications, the reconstruction performance will decline significantly. This is a common and important problem in clinical practice because different scanning requirements will lead to changes in voxel size. To solve this problem, the author proposes and studies several resolution - robust 3D MRI reconstruction methods based on 2D diffusion priors. These methods include: 1. **Data - driven method**: Train the diffusion model at multiple resolutions to make it robust to different resolutions. 2. **Interpolation method**: Use continuous representations (such as implicit neural networks and Gaussian interpolation) to alleviate the sensitivity caused by resolution changes. 3. **Infinite - resolution diffusion model**: Learn an image representation independent of resolution to meet the reconstruction requirements at different resolutions. Finally, the author finds that the **data - driven method of training the diffusion model at multiple resolutions** can effectively improve the reconstruction quality without sacrificing accuracy. In contrast, other model - based methods (such as Gaussian interpolation, neural representation, etc.) fail to significantly improve the performance degradation caused by resolution changes. ### Summary of key issues - **Limitations of existing methods**: Existing 3D MRI reconstruction methods based on diffusion models perform well at a fixed resolution, but their performance will decline significantly at different resolutions. - **Objective**: Develop a 3D MRI method that can maintain high - quality reconstruction at different resolutions. - **Solution**: Train the diffusion model at multiple resolutions to make the reconstruction method robust to resolution changes. ### Related formulas The formulas involved in this paper are mainly used to describe the training process of the diffusion model and the variational optimization method. For example: - The forward process of the diffusion model: \[ p(x_{t + 1}|x_t)=\mathcal{N}(x_{t + 1};\sqrt{1-\beta_t}x_t,\beta_tI) \] where \(\beta_t\) is the noise level. - The objective function of variational optimization: \[ \hat{x}(y)=\arg\min_{x\in\mathbb{C}^{W\times H\times D}}\sum_{i = 1}^C\|y_i - MF_3DS_ix\|^2_2+\lambda\mathbb{E}_{s\sim2D - Slices(x)}[R_{\text{diff}}(s)] \] where \(R_{\text{diff}}(s)\) is the regularization term based on the pre - trained 2D diffusion model, and \(\lambda\) is a hyperparameter. Through these methods, the author has successfully improved the robustness and accuracy of 3D MRI reconstruction at different resolutions.

Resolution-Robust 3D MRI Reconstruction with 2D Diffusion Priors: Diverse-Resolution Training Outperforms Interpolation

High-resolution MRI synthesis using a data-driven framework with denoising diffusion probabilistic modeling

InverseSR: 3D Brain MRI Super-Resolution Using a Latent Diffusion Model

Improving 3D Imaging with Pre-Trained Perpendicular 2D Diffusion Models

Using Diffusion Model to Generate High-Resolution MRI

Robust Physics-based Deep MRI Reconstruction Via Diffusion Purification

3D‐SRDM: 3D super‐resolution of MRI volumes based on diffusion model

Arbitrary scale super-resolution diffusion model for brain MRI images

MRI Reconstruction with Side Information using Diffusion Models

Rethinking Diffusion Model for Multi-Contrast MRI Super-Resolution

Joint Non-Linear MRI Inversion with Diffusion Priors

Beyond Nyquist: A Comparative Analysis of 3D Deep Learning Models Enhancing MRI Resolution

Model-Based Deep Learning for Reconstruction of Joint k-q Under-sampled High Resolution Diffusion MRI

Submillimeter diffusion MRI using an in-plane segmented 3D multi-slab acquisition and denoiser-regularized reconstruction

Reference-free Axial Super-resolution of 3D Microscopy Images using Implicit Neural Representation with a 2D Diffusion Prior

Physics-informed Deep Diffusion MRI Reconstruction with Synthetic Data: Break Training Data Bottleneck in Artificial Intelligence

INFusion: Diffusion Regularized Implicit Neural Representations for 2D and 3D accelerated MRI reconstruction

Slice-Consistent 3D Volumetric Brain CT-to-MRI Translation with 2D Brownian Bridge Diffusion Model

Diffusion Modeling with Domain-conditioned Prior Guidance for Accelerated MRI and qMRI Reconstruction

Robustness of Deep Learning for Accelerated MRI: Benefits of Diverse Training Data

Assessment of resolution and noise in magnetic resonance images reconstructed by data driven approaches