Answers are grounded in this paper and cite their sources.
X-ray computed tomography can quantify pores inside a material, but noisy and blurred pore boundaries make measurements sensitive to the segmentation procedure. This work evaluates direct regression of pore volume, surface area proxy, sphericity, and elongation from a centered synthetic grayscale crop. We generated 180 independent sphere, ellipsoid, and irregular pore geometries and rendered each under three degradation conditions, yielding 540 crops of size 24 3 . All renderings of each geometry remained in one split. A compact three-dimensional convolutional neural network (CNN) was compared with Otsu thresholding and an Extra Trees regressor using segmentation and intensity features. On 27 held-out geometries at high degradation, volume mean absolute percentage error was 14.76% for the CNN, 67.29% for Otsu, and 3.74% for Extra Trees. Otsu recovered clean discrete volume exactly and had the lowest pooled elongation error. Paired and out-of-domain experiments reveal how size, nearby pores, unresolved features, and sampling affect the results. The contribution is a reproducible synthetic benchmark that identifies conditions under which learned estimation improves on thresholding and conditions where classical measurements remain stronger. Independent industrial scans are needed to assess transfer.
No takes yet. Share an insight, caveat, or question.
Yadav et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: