Trueness of artificial intelligence-driven CBCT tooth segmentation: A comparative validation ex vivo pilot study

To evaluate and compare the trueness of four commercial AI-driven segmentation tools (Diagnocat, Relu, CephX, CoDiagnostiX) and one open-source software (3D Slicer with DentalSegmentator) for dental cone-beam computed tomography (CBCT) segmentation using extracted human teeth as physical ground truth reference.

Ten single-rooted teeth from two formaldehyde-fixed human donor heads were scanned using high-resolution CBCT (90 µm voxel size). Following CBCT acquisition, the teeth were extracted and digitized using an intraoral scanner (TRIOS 4) to create ground-truth reference models. DICOM datasets were processed by each segmentation software. 3D models were registered to reference scans of extracted teeth using landmark-based alignment followed by iterative closest point refinement. Surface-to-surface distances were calculated, including Hausdorff distance, root mean square distance (RMSD), mean absolute distance (MAD), median distance, and 95th percentile distance.

Significant differences were observed between segmentation tools for all metrics (Friedman test, p < 0.001). Relu demonstrated the highest trueness with Mean Distance of 0.11 mm [0.09-0.11], followed by Diagnocat 0.10 mm [0.09-0.13], CoDiagnostiX 0.14 mm [0.12-0.16], 3D Slicer 0.13 mm [0.12-0.15], and CephX 0.24 mm [0.21-0.25]. Post-hoc analysis revealed significant differences between top-performing (Relu, Diagnocat) and lower-performing tools (CephX). 3D Slicer showed performance similar to that of commercial solutions.

Relu and Diagnocat achieved the highest trueness for CBCT tooth segmentation when validated against physical specimens. All tested tools demonstrated clinically acceptable trueness for most applications, though performance varied considerably between platforms.

This pilot study provides preliminary comparative data on AI segmentation tool trueness that may inform future larger-scale validation studies. The findings suggest performance differences between platforms that warrant further investigation before definitive clinical recommendations can be made.