Abstract
<title>Abstract</title> <p>We developed and validated a deep learning model for the automated, simultaneous (multilabel) detection of artifacts and positioning errors in panoramic radiographs. A total of 1,500 retrospectively collected panoramic radiographs were annotated by a single experienced oral and maxillofacial radiologist (10 + years of clinical experience) for the binary presence of artifacts and positioning errors. Intra-observer reliability was assessed on a subset of 150 images after a 15-day interval using Cohen's kappa. The dataset was split using stratified sampling into training (n = 1,050; 70%), validation (n = 225; 15%), and test (n = 225; 15%) sets. An EfficientNet-B0 convolutional neural network pre-trained on ImageNet was fine-tuned within a multilabel classification framework using weighted binary cross-entropy loss. On the independent test set, the artifact detection model achieved an AUC-ROC of 0.839 (95% CI: 0.78–0.90), sensitivity of 0.721, specificity of 0.833, F1-score of 0.810, and precision of 0.923. The positioning error detection model achieved an AUC-ROC of 0.774 (95% CI: 0.70–0.85), sensitivity of 0.665, specificity of 0.696, F1-score of 0.763, and precision of 0.895. The proposed model demonstrated good diagnostic performance for artifact detection and moderate performance for positioning error detection in a single-center retrospective dataset, providing preliminary evidence supporting the feasibility of AI-assisted quality control in dental radiology workflows, although multicenter validation is required before clinical translation.</p>