Artificial intelligence-assisted identification of condensing osteitis and idiopathic osteosclerosis on panoramic radiographs

Abstract

Idiopathic osteosclerosis (IOS) and condensing osteitis (CO) represent radiopaque lesions often detected incidentally within the jaws, posing substantial diagnostic challenges due to their overlapping radiographic characteristics. The objective of this study was to assess the diagnostic efficacy of YOLOv8 and YOLOv11 deep learning algorithms in the identification of IOS and CO lesions on panoramic radiographs. A comprehensive collection of 1,000 panoramic images was retrospectively gathered and meticulously annotated utilizing a bounding box approach by two proficient oral and maxillofacial radiologists. All images were standardized to a resolution of 640 × 640 pixels and segregated into training (70%), validation (15%), and testing (15%) subsets. The performance of the models was evaluated based on metrics including accuracy, sensitivity, precision, F1 score, and the area under the receiver operating characteristic curve (AUC). YOLOv11 achieved notable precision scores of 98.8% for IOS and 97.1% for CO, alongside F1 scores of 96.8% and 95.6%, respectively. Conversely, YOLOv8 produced precision scores of 96.6% for IOS and 91.4% for CO, with F1 scores of 94% and 90%. These findings illustrate that AI-enhanced deep learning models possess the capability to accurately identify IOS and CO lesions, thereby presenting opportunities to improve diagnostic consistency, avert unnecessary invasive procedures, and facilitate more effective treatment planning within clinical practice.

Keywords: Condensing osteitis; Deep learning algorithms; Idiopathic osteosclerosis; Panoramic radiography.

Fig. 1

Annotated panoramic radiographs illustrating IOS, CO, and their coexistence, annotated using the CVAT tool. The middle column presents cases of CO, while the right column displays cases exhibiting both IOS and CO. Green bounding boxes indicate IOS lesions, whereas red bounding boxes highlight CO lesions.

Fig. 2

Correlogram showing the relationships between dataset classes, highlighting correlations and potential label dependencies. Darker shades represent stronger correlations, aiding in understanding data distribution challenges.

Fig. 3

Learning curves for training and validation for Yolo v11, including box loss, classification loss, and DFL loss. Also shown are trends for Precision, Recall, mAP@50, and mAP@50–95, indicating model convergence and overfitting prevention.

Fig. 4

Recall-confidence curves for both classes, illustrating how recall varies with confidence thresholds, reflecting the model’s ability to maintain high recall across different confidence levels.

Fig. 5

Precision-confidence curves for each class, showing the balance between precision and confidence thresholds, highlighting the model’s ability to minimize false positives.

Fig. 6

Sample predictions for Yolo v11 from the validation set, comparing model outputs to ground-truth labels. The images demonstrate the model’s ability to accurately detect and classify objects. The first images represent the ground-truth labels while the second images depict the predicted labels.

Fig. 7

Confusion matrix of the YOLOv8 model on the validation dataset. The matrix illustrates the distribution of true positive, false positive, false negative, and true negative predictions, offering a detailed assessment of the model’s detection accuracy. Notable misclassifications are observed in certain categories, indicating areas for further optimization.

Fig. 8

Confusion matrix of the YOLOv11 model derived from the validation dataset. This matrix provides an overview of the model’s predictive performance by visualizing class-wise agreement between predicted and ground-truth labels. The results highlight the model’s strengths in accurate detections while identifying specific classes where false predictions are more frequent.