What question did this study set out to answer?

To develop an automated system for assessing the quality of periapical radiographs using deep learning.

January 14, 2026Open Access

Deep learning-based assessment of periapical radiographic image quality

Key Points

To develop an automated system for assessing the quality of periapical radiographs using deep learning.
Utilized a retrospective dataset of 3594 periapical radiographs
Employed expert annotators for labeling images on tooth positions and quality defects
Trained seven ResNet50 models for classification and defect detection
Implemented data augmentation and oversampling to manage class imbalance
Evaluated model performance using metrics like AUC, accuracy, precision, recall, and F1 score.
Achieved AUC of 0.997 for tooth position classification
Rendered AUC values of 1.000 for several quality defects such as horizontal angle and crown coverage
Obtained high accuracy rates for all detected quality defects
Indicated potential for reducing retake rates and improving workflow efficiency in dental practices

Abstract

The manual assessment of periapical radiographs (PAs) for image quality is inherently subjective, labor-intensive, and time-consuming. This leads to diagnostic uncertainty, unnecessary patient radiation from retakes, and a lack of systematic quality improvement. To address this, we developed an automated deep learning system for comprehensive PAs quality assessment. A retrospective dataset of 3594 PAs was utilized. Expert annotators labeled each image for ten clinically relevant tooth-position classes and six common quality defects: vertical angle, horizontal angle, crown coverage, apical coverage, cone cut, and scratch. We trained seven independent ResNet50 models—one for multi-class tooth position classification and six for binary defect detection. To enhance robustness and mitigate class imbalance, extensive data augmentation and oversampling strategies were implemented during training. Model performance was comprehensively evaluated using standard metrics, including the area under the receiver operating characteristic curve (AUC), accuracy, precision, recall, F1 score, and loss value. The models demonstrated excellent performance, achieving AUC values of 0.997 for tooth position, and 0.996 (poor vertical angle), 1.000 (poor horizontal angle), 1.000 (incomplete crown coverage), 0.994 (incomplete apical coverage), 0.999 (cone cut), and 0.924 (scratch) for the various quality defects. These results indicate that the ResNet50 algorithm provides an effective and highly accurate approach for the automated detection of image quality issues in PAs. This AI tool holds significant potential for clinical translation, its integration into digital dental imaging systems could offer real-time, objective feedback, thereby improving diagnostic accuracy, reducing retake rates to minimize patient radiation exposure, and enhancing overall workflow efficiency in dental clinics. Validation on independent, multi-center datasets is required before clinical deployment.

Read Full Paperexternally

Mark Helpful

Bookmark

Relay

View Full Paper