Answers are grounded in this paper and cite their sources.
After a candidate defect has been located in an X-ray computed tomography (CT) volume, an inspector must still decide whether it is a pore, a crack or an inclusion, because each type has different acceptance criteria and root causes. Pores and cracks are voids with near-zero attenuation, whereas inclusions have a distinct, non-zero attenuation, so intensity should help where shape alone is ambiguous. We test this in a controlled synthetic study: 7,500 exactly labelled 3D patches (2,500 per class) generated from 200 simulated scans with partial-volume blur, shading and Poisson–Gaussian noise at three SNR levels, split by scan into training, validation and test sets. A compact 3D CNN (0.88 M parameters) reaches 0.990 test accuracy (0.990 ± 0.002 over five training seeds; 95% CI of the headline run 0.984–0.996). An identical network given only binary masks reaches 0.824 ± 0.015, a drop of 16.6 ± 1.6 points. In the headline run (drop 17.5 points; McNemar exact p ≈ 10⁻⁵⁰) the loss is confined to pore–inclusion confusion, cracks being recognised perfectly, and comes almost entirely from spherical inclusions, whose masks resemble spherical pores. Random Forest and SVM classifiers on shape and intensity features perform similarly (0.988–0.997), so the CNN is competitive rather than superior. All 11 errors of the headline run are either dark inclusions with contrast close to that of a void or pores and cracks of 59–69 voxels, and accuracy did not degrade from SNR 40 to 10. Because the data are synthetic, the results are an upper bound for real scans.
No takes yet. Share an insight, caveat, or question.
Rahul et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: