BigReg: An Efficient Registration Pipeline for High-Resolution X-Ray and Light-Sheet Fluorescence Microscopy

TL;DR

BigReg tackles the challenging problem of registering high-resolution XRM and LSFM volumes by combining a fast surface-feature stage with a Fourier-domain refinement. The two-stage pipeline first establishes a coarse, global alignment via surface point clouds (FPFH+RANSAC followed by point-to-plane ICP) and then performs a precise volumetric registration using masked normalized cross-correlation in the Fourier domain, yielding $T_{ ext{overall}}=T_2T_1$. On seven ex vivo mouse tibia pairs with landmark-based ground truth, BigReg achieves micrometer-scale accuracy ($LMD=8.36$ µm, LM fitness=85.71%), which improves to $LMD=7.24$ µm and LM fitness=93.90% when used as an initialization for mutual information-based methods. The approach is memory- and compute-efficient, completes in about ten minutes, and enables robust alignment of microscale bone structures like lacunae and osteocytes, advancing osteoporosis research through integrated multi-modal imaging.

Abstract

Recently, X-ray microscopy (XRM) and light-sheet fluorescence microscopy (LSFM) have emerged as pivotal tools in preclinical research, particularly for studying bone remodeling diseases such as osteoporosis. These modalities offer micrometer-level resolution, and their integration allows for a complementary examination of bone microstructures which is essential for analyzing functional changes. However, registering high-resolution volumes from these independently scanned modalities poses substantial challenges, especially in real-world and reference-free scenarios. This paper presents BigReg, a fast, two-stage pipeline designed for large-volume registration of XRM and LSFM data. The first stage involves extracting surface features and applying two successive point cloud-based methods for coarse alignment. The subsequent stage refines this alignment using a modified cross-correlation technique, achieving precise volumetric registration. Evaluations using expert-annotated landmarks and augmented test data demonstrate that BigReg approaches the accuracy of landmark-based registration with a landmark distance (LMD) of 8.36\,\textmu m\,$\pm$\,0.12\,\textmu m and a landmark fitness (LM fitness) of 85.71\%\,$\pm$\,1.02\%. Moreover, BigReg can provide an optimal initialization for mutual information-based methods which otherwise fail independently, further reducing LMD to 7.24\,\textmu m\,$\pm$\,0.11\,\textmu m and increasing LM fitness to 93.90\%\,$\pm$\,0.77\%. Ultimately, key microstructures, notably lacunae in XRM and bone cells in LSFM, are accurately aligned, enabling unprecedented insights into the pathology of osteoporosis.

Figures (5)

• Figure 1: Visualization of (a) a mouse tibia sample used as the scanning subject and 3D rendered volume for (b) XRM data and (c) LSFM data. Both volumes are grayscale but shown in different colors, and their voxel sizes and shapes are unified to 1.42 $\times$ 1.42 $\times$ 1.42 µ m$^3$ and 2048 $\times$ 2048 $\times$ 800, respectively. While the XRM scan reconstructs the complete bone structure, the LSFM scan omits the lower part of the bone cross-section due to increased scattering along the detection depth. In the highlighted subregions within the orange boxes, which are magnified threefold, the XRM volume features vessel canals (blue labels) and lacunae (red labels), whereas the LSFM volume shows vessels (blue labels) and cell nuclei (red labels).
• Figure 2: Illustration of BigReg pipeline. From left to right, the moving XRM volume needs to be rigidly transformed to align with the fixed LSFM volume. Stage 1 extracts surface features and performs two successive point cloud-based registration procedures to offer a coarse alignment $\textbf{T}_1$. Stage 2 refines this alignment using masked normalized cross-correlation (MNCC) in the Fourier domain to achieve finer volumetric registration $\textbf{T}_2$. Finally, the registration result is automatically obtained by combining these two transformations.
• Figure 3: Illustration of scanning procedure of (a) XRM and (b) LSFM. Specifically, XRM captures high-resolution projection images through the two-stage magnification architecture; LSFM records horizontal slices illuminated by the corresponding thin light sheet.
• Figure 4: Visualization of the registration process in BigReg. Rows display central slices from different planes of the same sample. Columns one and two show the entire views at the initial state and S1.1, respectively, with specified dimensions. The subsequent three columns highlight eightfold magnified regions of interest (ROIs) from the $256\times256$ orange boxes through successive registration stages. The last column displays the GT result.
• Figure 5: Qualitative results on an extremely misregistered example. Rows display central slices from different planes of the same sample. The first column shows the initial state in full view. Except for the second column which presents MMI's failed result, subsequent columns highlight eightfold magnified ROI from the $256\times256$ orange boxes, corresponding to the applied methods.