Using of artificial intelligence to evaluate image quality for interproximal bone level assessment on bitewing...

To develop an artificial intelligence (AI) model capable of certifying whether the quality of the bitewing image is adequate or compromised, as an intermediate tool for the technical analysis prior to the final evaluation of the alveolar bone crest.

A total of 372 bitewing radiographs from a radiographic database were classified by three evaluators into control 186 (standard vertical angulation) and test 186 (offset vertical angulation) groups, forming a consensus gold standard for accuracy with the algorithm classification. The region of interest was defined as the non-overlapping area between the buccal and palatal/lingual cusps for data labeling, using the YOLOv8n-Seg architecture for training, validation, and testing. The model's performance was evaluated by comparing the average areas among groups, computing the metrics of precision, sensitivity, specificity, accuracy, recall and F1 score.

The AI model achieved a classification accuracy of 96.7%, with 100% sensitivity and 93.8% specificity (ICC = 0.95). The precision was 0.93, the recall 1.0 and the F1-score 0.96. Across both training and validation stages, significant differences were observed between the test and control groups (P < 0.05), with consistently larger regions of interest areas in the test group. A direct relationship was observed between bitewing radiographs classified as having offset vertical angulation and the regions of interest area. The model achieved excellent accuracy, comparable to that of expert evaluators.

The AI model accurately segmented and classified whether or not vertical angulation was compromising the quality of bitewing radiographs, thereby minimizing errors in the diagnostic interpretation of the alveolar bone crest.