Caries Detection in Primary Molars with Bitewing Radiographs through Deep Learning Based-Object Detectors

Automated methods for caries detection among children can help triage children with greater need for treatment and assist clinicians in achieving a more accurate diagnosis of caries lesions. Therefore, we aimed to develop artificial intelligence algorithms based on various object detectors for detecting and staging caries lesions in primary molars using bitewing radiographs.

We used a dataset of 1,023 bitewing radiographs of primary molars from children aged 3-10 years. The radiographs were annotated by two examiners using a dedicated annotation platform to draw bounding boxes corresponding to four different primary caries severity stages. Five popular deep learning object detection algorithms were trained to detect the annotated caries and evaluated with various performance metrics, considering three different thresholds: all lesions, dentine caries lesions, and dentine caries lesions requiring operative treatment.

For staging caries lesions, the DINO model achieved higher concordance with a weighted kappa score of 0.713 on the test dataset, outperforming other object detectors. The DINO model achieved the highest sensitivity for detecting all caries lesions (0.509) while also attaining the highest sensitivity and accuracy in identifying dentine caries requiring operative treatment, at 0.659 and 0.980, respectively. However, YOLOv7 also achieved good performance, with specificity values exceeding 0.99 and accuracy values exceeding 0.92 across all thresholds.

The DINO and YOLOv7 algorithms perform well at detecting caries on primary molars in bitewings, highlighting their potential for clinical application at aiding clinicians in daily practice.