Radiographic assessment of post-endodontic filling features on PAN and CBCT: diagnostic agreement of an AI platform against CBCT consensus

This retrospective diagnostic accuracy study evaluated the performance of an artificial intelligence (AI) platform (Diagnocat) in assessing endodontic treatment features via panoramic (PAN) and cone-beam computed tomography (CBCT) images from 163 patients. Two experienced observers (a radiologist and a dentist) provided consensus readings on the CBCT images, which served as the reference standard. Because the platform applies modality-specific processing pipelines, analyses were conducted and reported separately for CBCT and PAN. The AI analyzed five treatment variables-adequate obturation, adequate density, overfilling, voids in filling, and short filling-and its performance was compared against the reference standard for both the PAN and the CBCT. Diagnostic accuracy, precision, recall (sensitivity), and F1-scores were calculated. Diagnocat exhibited excellent diagnostic performance on CBCT images, achieving overall accuracy above 94% and perfect (100%) sensitivity for overfilling. On PAN benchmarked against the CBCT reference, performance was lower (accuracies 68.25-84.66%), with limited precision and F1-scores for adequate obturation and adequate density, while sensitivity remained comparatively high for voids and short fillings. These results demonstrate modality-dependent platform performance under the studied acquisition protocols and reference standard.