Influence of Geometry, Class Imbalance and Alignment on Phantom Reconstruction Accuracy - A Micro-CT Phantom-Based Evaluation

The accuracy of the 3D models created from medical scans depends on every stage of the pipeline - segmentation, alignment, and surface extraction. Although segmentation algorithms are well established, the effects of the type of geometry, class imbalance and voxel and point cloud alignment on accuracy remain to be sufficiently explored. This work evaluates the reconstruction accuracy and also studies the effect of voxel alignment on overlap accuracy metrics using controlled perturbation analysis. A sphere, a facemask, and an abdominal aortic aneurysm (AAA) were printed using the stereolithography technique and scanned using a micro-CT machine. Segmentation was performed using Gaussian Mixture Model (GMM), Otsu thresholding and region-growing (RG) based methods. Segmented and reference models were aligned using the Kabsch-Umeyama (KU) algorithm to evaluate Dice and Jaccard scores, precision, among others. A sensitivity analysis quantified the effect of controlled translational (±5 voxels) and rotational (±5°) perturbations on post-KU-aligned voxel grids. Surface meshes, generated using the marching cubes algorithm, were registered with reference meshes using RANSAC and ICP alignment process. Otsu thresholding yielded highest accuracy for all the geometries. The sphere, with uniform curvature and no thin-walled regions, showed the lowest sensitivity to alignment error. AAA yielded low overlap scores due to its small wall thickness and sensitivity to misalignment. AAA model was most sensitive to misalignment, with Dice and Jaccard scores decreasing by more than 90% under 5° rotation and by over 60% due to just 1°. The facemask surface was most error-prone, possibly due to misalignment during the ICP process. The Jaccard index yields numerically lower values and larger relative changes under perturbations than the Dice score, making alignment sensitivity more apparent for thin-walled structures. Even a sub-millimetre error in voxel alignment adversely affects accuracy. Surface-based accuracy metrics differed from the voxel-based trends, and this divergence carries diagnostic value.