Efficient and robust 3D blind harmonization for large domain gaps

TL;DR

The paper tackles MRI domain gaps by proposing BlindHarmonyDiff, a 3D blind harmonization framework that uses a target-domain edge-to-image model to harmonize unseen source-domain MR images. It introduces a rectified-flow-based E2I generator trained with multi-stride patch training and a refinement module to suppress hallucinations and enforce cross-domain consistency. The core loss is defined over a velocity field $v_\theta$ mapping Gaussian noise $x_0$ to edge-conditioned patches $x_1$, optimizing $\mathcal{L}= \mathbb{E}_{x_0\sim\mathcal{N}(0,1)} [ \int_0^1 \lVert (x_1 - x_0) - v_\theta (x_t,t; e,i,j,k) \rVert^2 dt ]$, enabling high-quality 3D harmonization. Experiments on the OASIS3 dataset across multiple scanners show improved PSNR/SSIM and downstream task performance, demonstrating robustness and practical applicability for blind harmonization in MRI pipelines.

Abstract

Blind harmonization has emerged as a promising technique for MR image harmonization to achieve scale-invariant representations, requiring only target domain data (i.e., no source domain data necessary). However, existing methods face limitations such as inter-slice heterogeneity in 3D, moderate image quality, and limited performance for a large domain gap. To address these challenges, we introduce BlindHarmonyDiff, a novel blind 3D harmonization framework that leverages an edge-to-image model tailored specifically to harmonization. Our framework employs a 3D rectified flow trained on target domain images to reconstruct the original image from an edge map, then yielding a harmonized image from the edge of a source domain image. We propose multi-stride patch training for efficient 3D training and a refinement module for robust inference by suppressing hallucination. Extensive experiments demonstrate that BlindHarmonyDiff outperforms prior arts by harmonizing diverse source domain images to the target domain, achieving higher correspondence to the target domain characteristics. Downstream task-based quality assessments such as tissue segmentation and age prediction on diverse MR scanners further confirm the effectiveness of our approach and demonstrate the capability of our robust and generalizable blind harmonization.

Figures (10)

• Figure 1: (a) Previous blind harmonization methods have limitations such as 2D reconstruction, moderate image quality, and limited performance for a large domain gap. (b) BlindHarmonyDiff is a 3D blind harmonization method designed to overcome the limitations of the previous methods. BlindHarmonyDiff trains an edge-to-image model only with the target domain image and generates a harmonized image by inputting an edge map from the source domain image into the trained model. A refinement module then enhances the harmonized image to maintain a high correlation with the input source domain image. (c) An edge-to-image model is trained with a newly proposed multi-stride patch processing. As compared to a previous patch method, which simply creates a patch from neighboring voxels of an image, a multi-stride patch method randomly selects strides to construct a patch.
• Figure 2: PSNR and SSIM across different multi-stride patch ratios. The ratio of multi-stride patches used in training ranged from 0% to 100% at 20% intervals.
• Figure 3: Visual comparison of harmonization results on traveling subjects, demonstrating the effectiveness of BlindHarmonyDiff against the other blind harmonization methods.
• Figure 4: Tissue segmentation results for the various harmonization methods, with the segmentation network trained on the target domain data. Each column represents a different harmonization approach applied to the input images.
• Figure 5: Comparison of 3D vs. 2D image processing networks. 2D processing shows inter-slice discontinuities (see zoomed-in image), while 3D processing yields homogeneous results across slices.