Surface Reconstruction for Computed Tomography Volumes

Abstract

In the field of computed tomography, surface extraction from volumetric images is a fundamental capability with applications in both medical and industrial contexts. One particular application in the industrial field is metrology, the science of measurements, a discipline that imposes particularly strict accuracy requirements.

This thesis investigates surface extraction techniques, focusing on a pipeline that first estimates a rough surface and then refines it using sub-voxel methods. We evaluated several algorithms from the open-source library Visual Toolkit by comparing the extracted surfaces to a reference model and assessing whether it satisfies some reasonable metrology criterion. These experiments were conducted on CT data acquired using both circular and helical scanning geometries to verify the hypothesis that more accurately reconstructed input data would yield more accurately extracted surfaces. While helical scans improved the quality of the extracted surface, the results were not accurate enough to satisfy the metrology criterion.

We then implemented and analyzed two sub-voxelic refinement techniques: a method based on the center of mass of the image values and a method based on the gradient of the image values. The latter showed greater potential, leading us to conduct a study of its parameters and design choices, including the method of gradient computation, interpolation methods, and the method used to estimate the orientation of the surface normal. With an optimized configuration and high quality of input data, our method produced surfaces where 60% of the points met the metrology criterion, and the mean error also remained below the specified threshold.

Although we did not manage to achieve full compliance, our results indicate promise for applications such as the detection of porosities or the ability to get a precise slice view of the object. We also got a better understanding of the effect of our design choices and determined that future research should be focused on accurately reconstructing sharp features.