Abstract
Abstract
Objectives
To develop a deep learning model capable of automatically detecting common positioning errors in panoramic radiographs and to establish an error-based objective 10-point image quality scoring system aligned with expert evaluation.
Methods
A total of 4390 panoramic radiographs were retrospectively annotated for 7 clinically relevant positioning errors (including chin position, head rotation/tilt, bite position, tongue position, cervical spine positioning, and anatomical coverage) associated with diagnostically meaningful geometric distortion or anatomical superimposition by 2 experienced oral radiologists. A multi-label deep learning model based on a modified residual network was trained to detect all errors simultaneously. An independent dataset of 510 radiographs was scored for subjective image quality on a 10-point scale. Multiple linear regression was used to estimate the independent contribution of each positioning error to subjective image quality and to derive an objective scoring system established by converting regression-derived weights of significant errors into standardized point deductions.
Results
High annotation and scoring reliability were achieved (Kappa: 0.879-0.953; ICC: 0.838). The model achieved high performance across all error types, with accuracy ranging from 88.40% to 96.00% and area under the curve values between 98.17 and 99.37. Several positioning errors were associated with lower subjective image quality scores (P
Conclusions
The proposed system enables accurate detection of positioning errors and provides a structured and interpretable, error-based surrogate measure of perceived panoramic image quality.
Advances in knowledge
This study presents an error-specific, data-driven framework for objective assessment of panoramic image acquisition quality, enhancing the interpretability and standardization of image quality evaluation.