Automated Detection of Positioning Errors in Bitewing Radiographs Using Deep Learning.
Authors
Affiliations (3)
Affiliations (3)
- Dentomaxillofacial Radiology, Sydney Dental School, Faculty of Medicine and Health, The University of Sydney, Camperdown, New South Wales, Australia.
- School of Computer Science, Faculty of Engineering, The University of Sydney, Camperdown, New South Wales, Australia.
- Restorative and Reconstructive Dentistry, Sydney Dental School, Faculty of Medicine and Health, The University of Sydney, Camperdown, New South Wales, Australia.
Abstract
Bitewing radiographs (BWs) are essential diagnostic tools in dentistry, but are often affected by positioning errors, compromising their diagnostic value and increasing retake rates. This study aimed to develop and evaluate a deep learning-based automated system for detecting and classifying BW type (premolar/molar), side (right/left), and common positioning errors in a controlled mannequin-based setting. A total of 403 BWs were collected from dental student assessments acquired from dental radiology mannequins. BWs were consensus-labeled by dentomaxillofacial radiology experts for type, side, and the most common positioning errors. A balanced data set supported training and validation. Convolutional neural network architectures, including ResNet and DenseNet variants, were trained from scratch and evaluated as independent classification tasks using stratified fivefold cross-validation. Performance was assessed using accuracy, precision, recall, F1 score, and area under the curve, reported as mean ± standard deviation across folds. A confusion matrix was generated for multi-class cone-cutting error (CCE) severity classification. The system achieved accuracies of 95.8% for BW type classification, 99.0% for side classification, 96.3% for CCE detection, and 93.4% for interproximal overlap detection. Multi-class CCE classification achieved an accuracy of 79.3%, with reliable identification of "critical" and "minimal" errors. Detection of incorrect receptor placement achieved an accuracy of 73.2%. The CNN-based system showed strong performance for bitewing type, side, binary cone-cutting error, and interproximal overlap classification in mannequin-acquired radiographs. Performance was more limited for multi-class cone-cutting severity and incorrect receptor placement. These findings support the potential use of deep learning for radiographic quality assessment in education and quality-control workflows. A validation study using patient-acquired bitewing radiographs from clinical settings is required before clinical deployment.