In this research project we register 2D petrographic light microscopy (LM) thin sections into 3D micro-CT volumes of freeze-thaw damaged concrete specimens, using deep learning segmentation models (i.e., Residual Attention UNets) for pixelwise classification of cracks, air pores, aggregates, cement matrix followed by 3D Fourier-based spatial-harmonic feature matching rather than image-intensity registration. Among other two 2D registration workflows tested, only the 3D method gave sufficiently accurate and visually acceptable results with 89.75% success rate of convergence to a solution and 5.21% minimum relative RMSE.
A petrographic thin section is a physical slice cut from the specimen at an arbitrary angle, so the task is to find which plane inside the 3D CT stack corresponds to the 2D LM slice with an arbitrary different isotropic scaling magnitude. The method must be invariant to translation, rotation, reflection, and scaling, because the LM and CT scans differ in pixel resolution (0.85 µm vs. ~20.88 µm). Registration inputs for the published case study are the segmentation masks of both scanning modalities rather than raw images. The pipeline comprises four stages, namely canonical normalization of 3D shape instances, spherical projection and spatial harmonics, feature matching, and plane fitting for inverse transformation estimation.
First, for a point set P = {p1, …, pk}, the shape is centered by subtracting its centroid c, then oriented via eigen-decomposition of the covariance matrix, which gives orthonormal eigenvectors forming the rotation matrix R. Reflection ambiguity in the eigenvector directions is resolved with a diagonal sign matrix
F = diag(sign(fx), sign(fy), sign(fz)),
built from the coordinate sums of cubes, and scale is normalized by the root-mean-square distance of the points from the origin, giving the canonical form
p̂ = τ(p) = s−1 F R (p − c).
Second, the normalised points are projected onto a unit sphere, triangulated (spherical Delaunay), and treated as a mesh. The Laplace–Beltrami operator is discretised with FEM (mass matrix B from element areas, stiffness matrix S from cotangent weights) and its generalized eigenproblem
S ul = λl B ul
is solved; the eigenvalues are spatial frequencies and the eigenvectors the harmonic basis functions. The scalar radial-distance function on the mesh is projected onto this basis, and the resulting coefficients c = [c1, …, cL] are rotation-, translation-, reflection-, and scaling-invariant descriptors. This is what replaces the ReResNet used by the adapted Brandstätter pipeline, which was scaling- and translation-dependent.
Third, descriptor vectors of keypoints on the source and target shapes are compared by Euclidean distance, and a pair is a match when
‖cS,i − cT,j‖ < ε.
Matching happens between the original keypoints (the descriptors only serve as the comparison medium), producing n matched point pairs. In the experiments, keypoints were drawn from three segmentation classes (crack junctions, aggregate boundaries, air-pore centroids) and their combinations.
Fourth, a plane is fitted to the matched source points in normalized space by PCA: the eigenvector of the smallest eigenvalue of the matched-point covariance matrix is the plane normal, giving plane norm. The plane is then mapped back to the original coordinate system by inverting the canonical transform: points transform as
p = s RT F p̂ + c,
while normals transform via the inverse transpose,
norig = s−1 RT F nnorm,
yielding the final plane
norig · (x − sorig) = 0
in the CT volume, i.e., the location of the LM slice.
Validated on 75 registration experiments over synthetically generated annotated volumes to remove segmentation errors as a confound, the 3D method reached 89.75% success with a minimum RRMSE of 5.21%, versus 56.68% and 43.29% for the fastest 2D-only cross-section method. The price is computation: it was the slowest of the three, sometimes exceeding 24 h per run on a 16-core Xeon with an RTX A5000, and about 45% of failed attempts converged on a second run with a different random seed. The authors note the method also depends on having accurate segmentation masks as input, which adds a trainability burden when devices, settings, or concrete types change.
Source Code & Supplementary Materials
- 1 - Github Repository: dfg-rissanalyse
- 2 - Zenodo: Petrographic Reflected Light Microscopy and Computed Tomography Scans of Concrete Specimens
Alabassy, Mohamed Said Helmy; Hampe, Martin Christian; Erfurt, Doreen; Ludwig, Horst-Michael; Osburg, Andrea (2026). [Dataset]. Zenodo. DOI: 10.5281/zenodo.19037136.
Publications
- Fourier-based Non-rigid Slice-to-volume Registration of Segmented Petrographic LM and CT Scans of Concrete Specimens
Alabassy, Mohamed Said Helmy; Hampe, Martin Christian; Erfurt, Doreen; Ludwig, Horst-Michael; Osburg, Andrea (2026). In: Materials, 19(4), 663, MDPI. DOI: 10.3390/ma19040663. - Automated segmentation of cracks in light microscopic images of concrete using machine learning
Sartori, Julian; Alabassy, Mohamed Said Helmy; Erfurt, Doreen; Osburg, Andrea; Ludwig, Horst-Michael (2025). In: 5th International Conference on the Chemistry of Construction Materials (5th ICCCM). Date: 08.09. - 10.09.2025. Garching, Germany. - Unsupervised Semantic Segmentation of Cracks in Concrete Specimens
Sadrnia, Amir; Alabassy, Mohamed Said Helmy; Osburg, Andrea (2025). In: Proceedings of the 36th Forum Bauinformatik (p. 90-97). Date: 24.09. - 26.09.2025. Aachen, Germany. DOI: 10.18154/RWTH-CONV-254881. - Combination and Automation of Computer-aided Methods for Quantitative Crack Analysis of Concrete
Patzelt, Max; Alabassy, Mohamed Said Helmy; Ludwig, Horst-Michael; Osburg, Andrea (2024). In: 7th International Conference on Concrete Repair, Rehabilitation and Retrofitting (ICCRRR). Date: 04.11. - 06.11.2024. Cape Town, South Africa.