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

TL;DR

The paper tackles the problem of robust 3D MRI reconstruction when voxel size varies in practice, proposing a variational framework that regularizes 3D reconstructions with 2D diffusion priors trained on slices from multiple resolutions. A key contribution is randomized slicing, which enforces 3D consistency while allowing the diffusion prior to operate on 2D planes; the regularizer uses a truncated diffusion objective with a stabilized Jacobian. The study systematically compares data-centric diverse-resolution training against model-centric interpolation and infinite-dimensional diffusion strategies, showing that diverse-resolution training most effectively mitigates resolution-induced artifacts without sacrificing accuracy. Practically, this approach enables more reliable 3D MRI reconstructions across scanner settings and clinical requirements, as demonstrated on knee and brain datasets, including ultra-high-field data.

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.

Figures (16)

• Figure 1: 3D reconstruction with 2D diffusion priors is sensitive to resolution-shifts. We reconstruct complex 3D volumes from undersampled multicoil 3D brain MRI measurements at a fixed voxel-size $V_{\text{rec}}$, and employ variational regularization with 2D diffusion-models pre-trained at voxel sizes with $V_{\text{train}} = V_{\text{rec}}$ (no shift), $V_{\text{train}} = 2 V_{\text{rec}}$ ($2 \times$-shift) and $V_{\text{train}} = 4 V_{\text{rec}}$ ($4 \times$-shift). We observe that differences in voxel size lead to significant artifacts in the reconstructed images, both visually (left panel) and quantitatively (right panel). As we show in our paper, training on diverse resolutions is an effective solution.
• Figure 2: Illustration of the proposed variational 3D MRI reconstruction method, which in each iteration regularizes randomly sampled slices with a diversely trained 2D diffusion prior.
• Figure 3: Diffusion-based and compressive sensing reconstruction of 3D MRI volumes. The right panel shows the PSNR of the reconstructions obtained from performing 3D reconstruction from undersampled multicoil 3D MRI measurements. It can be seen that the proposed variational method yields superior reconstruction quality, followed by posterior sampling and the compressive sensing baselines. Left panel: Slices from the reconstructed volumes demonstrate that the variational approach provides competitive visual quality and smoother reconstructions.
• Figure 4: Reconstruction under resolution shifts. 3D reconstruction on the Stanford 3D knee dataset at small, medium, and large voxel-sizes $V_{\text{rec}}$ with diffusion models trained at fixed or diverse voxel-sizes $V_{\text{train}}$. Performance drops when the test-voxel size is different to the train-voxel size, with larger gaps for higher accelerations and larger resolution-shifts. Training the model on diverse resolutions provides a robust reconstruction method across resolutions.
• Figure 5: Reconstruction at larger voxel-sizes ($V_{\text{rec}} > V_{\text{train}}$). Effectiveness of several volume and kernel interpolation methods, where the diffusion models are trained on $V_{\text{train}} = V_{2 \times}$, and we reconstruct at $V_{\text{rec}} = V_{4 \times}$. We observe that simple trilinear resampling mitigates the performance drop completely.