CBCT versus Machine-Learning Artificial Intelligence for Detecting and Characterizing Oral Lesions in Prosthodontic Patients

Abstract

Prosthodontic patients often have undetected oral hard- and soft-tissue lesions that can end up hidden beneath a prosthesis. Interpreting CBCT depends on the observer and takes time, so machine-learning–based AI may offer a useful alternative. Thus, the aim of our study was to interpret and compare the diagnostic efficacy of CBCT-based evaluation versus machine-learning–based AI analysis in the detection and characterization of oral lesions among prosthodontic patients among 100 prosthodontic patients with suspected oral lesions, who were randomly allocated into two groups of 50 each. In Group 1, two blinded, calibrated oral radiologists interpreted the CBCT scans. In Group 2, a validated convolutional neural network–based model analysed the imaging and clinical datasets. We have found that, AI showed higher sensitivity (91.67% vs 85.71%), specificity (85.71% vs 80.00%) and accuracy (90.00% vs 84.00%) than CBCT, with a significantly greater AUC (0.887 vs 0.829; p = 0.041). Agreement with the gold standard improved from κ = 0.68 to κ = 0.83. AI reduced interpretation time (1.24 vs 8.46 min; p < 0.001) and measurement deviation (1.08 vs 1.62 mm; p = 0.001), and was more reproducible (ICC 0.98 vs 0.86). Therefore we have found that, Machine-learning analysis is a rapid, reproducible and accurate adjunct to CBCT interpretation for pre-prosthetic assessment of oral lesions.