Accuracy of artificial intelligence-based detection of dental crowding and spacing on cone-beam computed tomography

Abstract Background To evaluate the accuracy of an artificial intelligence (AI) system for detecting anterior dental crowding and spacing on cone-beam computed tomography (CBCT) and to characterise clinically relevant patterns of diagnostic failure. Methods In this retrospective diagnostic accuracy study, CBCT scans from 88 patients were analysed. AI-generated diagnoses at the tooth-pair level were compared with a reference standard based on a modified Little’s Irregularity Index (mLII) that incorporates signed contact-point displacement. Diagnostic performance — sensitivity, specificity, positive and negative predictive values, and Cohen’s κ — was assessed at the tooth-pair level ( n = 875) across five thresholds (± 0.5 to ± 2.5 mm). Arch-level performance was assessed using the conventional Little’s Irregularity Index (LII). Statistical significance was set at α = 0.05, and all estimates are reported with 95% confidence intervals. Results At the tooth-pair level, sensitivity for crowding ranged from 28.8% to 42.0%, whereas specificity remained high (81.9–88.9%). Agreement with the reference standard was fair (maximum κ = 0.247). A consistent blind spot was identified at the maxillary central-incisor pair (11–21), where no crowded cases were detected despite confirmed irregularities. Spacing detection was poor (sensitivity 10–25%). At the arch level, detection improved with increasing severity of irregularity (38.8–75.0%) but remained incomplete and was accompanied by a high false-positive rate (41.2% in perfectly aligned arches). Conclusions The AI system showed limited sensitivity and only fair agreement for detecting anterior crowding and spacing on CBCT. Clinically relevant irregularities, including those at the anterior midline, may be missed. Because crowding and spacing are most reliably assessed from clinical examination and intraoral records rather than from CBCT, AI-based CBCT assessments should be interpreted with caution and cannot replace clinician evaluation.