Title: Aligning CAD Models with CT Scans of Cylinders Using Ellipse-Driven Axis Determination

Abstract: Establishing precise ground-truth geometry in volumetric imaging relies heavily on the accurate alignment of CAD models with CT scans. As deep learning architectures become increasingly sophisticated, the demand for large-scale datasets to train these models is rising, making the acquisition of reliable object masks critical. Conventional intensity-based techniques often falter when CT grayscale values are uncalibrated, while feature-based approaches like RANSAC or ICP struggle due to the lack of corresponding features between idealized CAD designs and noisy CT volumetric data. To address these challenges, we introduce a two-stage geometric registration framework tailored for cylindrical objects, such as ionization chambers, leveraging their specific geometric characteristics. The initial phase involves determining the 3D rotation axis by identifying elliptical cross-sections within CT slices. This process entails fitting ellipses to contours identified via edge detection, followed by Principal Component Analysis (PCA) on the centers of these ellipses after outliers have been eliminated using RANSAC. In the second phase, the CAD model is voxelized and aligned with the detected axis, with its position fine-tuned through translation to maximize volumetric overlap with the CT scan. This method delivers robust registration, maintaining tilt and orientation errors under $0.1^\circ$, without requiring intensity calibration or feature matching. The resulting aligned CAD model serves as a ground truth reference, facilitating applications such as automated analysis in industrial CT workflows and machine learning-based object localization.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC