Wattanapong Suttapak

Accurate classification of dental caries from bitewing radiographs is crucial for effective diagnosis and treatment planning. Traditional diagnostic methods can be subjective and inconsistent, highlighting the potential benefits of adopting computer-aided diagnostic systems. Recent studies have demonstrated the effectiveness of deep learning-based classification using the ICCMS^TM 7-class system, which is pivotal for precise caries diagnosis. Building on these advances, this study aims to enhance ResNet models by incorporating the projected gradient descent (PGD) into the training process. By introducing mild perturbations to augment the clean dataset, the proposed approach aims to enhance both the robustness of the model and its classification performance. The experimental results show that the PGD-augmented ResNet-50 exhibits the most significant performance improvements, with the validation accuracy increasing from 58.20 to 67.20% and the test accuracy improving from 57.14 to 59.18%. Additionally, sensitivity and specificity metrics showed notable gains, highlighting the clinical relevance of the approach. These results suggest that incorporating adversarial training techniques such as PGD can substantially improve the accuracy, robustness, and reliability of deep learning models for dental caries classification, thereby advancing computer-aided diagnostic tools in clinical practice.

Purpose To evaluate the performance of a deep learning-based convolutional neural network (CNN) algorithm and compare it with that of general dentists and dental students in classifying caries lesions in primary teeth using bitewing radiographs. Methods A total of 1400 bitewing radiographs (4715 tooth images) were divided into training, validation, and testing datasets, with carious lesions classified into four and seven classes. After training, the best-performing ResNet model was selected and compared with three general dentists and three dental students via a reference test. The accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), macro F1 score, area under the curve (AUC), and confusion matrices were evaluated. Results ResNet-152 outperformed ResNet-50 and ResNet-101 in the validation process. In the 4 class classification, ResNet-152 and Dental student 1 achieved accuracies exceeding 0.7, while most examiners ranged between 0.62 and 0.67. Only Dental student 1 and ResNet-152 achieved specificity and NPV values of 0.9 or higher. ResNet-152 and most examiners exhibited lower sensitivity for initial and moderate lesions than for extensive lesions. In the 7 class classification, the accuracy ranged from 0.37 to 0.58, with the best-performing comparators—Dental student 1, ResNet-152, and General dentist 3 exceeding 0.5. The sensitivity, PPV, and macro F1 score followed similar trends. ResNet-152 achieved a favourable AUC of 0.85. Conclusion ResNet-152 performed comparably to its leading human comparators, general dentists and dental students, demonstrating favourable performance in caries lesion classification. CNNs could serve as a second option in caries lesion classification, potentially leading to improved treatment decisions.

The study aimed to evaluate the impact of image size, area of detection (IoU) thresholds and confidence thresholds on the performance of the YOLO models in the detection of dental caries in bitewing radiographs. A total of 2575 bitewing radiographs were annotated with seven classes according to the ICCMS^™ radiographic scoring system. YOLOv3 and YOLOv7 models were employed with different configurations, and their performances were evaluated based on precision, recall, F1-score and mean average precision (mAP). Results showed that YOLOv7 with 640 × 640 pixel images exhibited significantly superior performance compared to YOLOv3 in terms of precision (0.557 vs. 0.268), F1-score (0.555 vs. 0.375) and mAP (0.562 vs. 0.458), while the recall was significantly lower (0.552 vs. 0.697). The following experiment found that the overall mAPs did not significantly differ between 640 × 640 pixel and 1280 × 1280 pixel images, for YOLOv7 with an IoU of 50% and a confidence threshold of 0.001 (p = 0.866). The last experiment revealed that the precision significantly increased from 0.570 to 0.593 for YOLOv7 with an IoU of 75% and a confidence threshold of 0.5, but the mean-recall significantly decreased and led to lower mAPs in both IoUs. In conclusion, YOLOv7 outperformed YOLOv3 in caries detection and increasing the image size did not enhance the model’s performance. Elevating the IoU from 50% to 75% and confidence threshold from 0.001 to 0.5 led to a reduction of the model’s performance, while simultaneously improving precision and reducing recall (minimizing false positives and negatives) for carious lesion detection in bitewing radiographs.

Objectives To assess the feasibility of the YOLOv3 model under the intersection over union (IoU) thresholds of 0.5 (IoU₅₀) and 0.75 (IoU₇₅) for caries detection in bitewing radiographs based on the International Caries Classification and Management System (ICCMS™). Materials and methods We trained the YOLOv3 model by feeding 994 annotated radiographs with the IoU₅₀ and IoU₇₅ thresholds. The testing procedure (n = 175) was subsequently conducted to evaluate the model’s prediction metrics on caries classification based on the ICCMS™ radiographic scoring system. Results Regarding the 4-class classification representing caries severity, YOLOv3 could accurately detect and classify enamel caries and initial dentin caries (class RA) (IoU₅₀ vs IoU₇₅: precision, 0.75 vs 0.71; recall, 0.67 vs 0.64). Concerning the 7-class classification signifying specific caries depth (class 0, healthy tooth; classes RA1-3, initial caries affecting outer half, inner half of enamel, and the outer 1/3 of dentin; class RB4, caries extending to the middle 1/3 of dentin; classes RC5-6, extensively cavitated caries affecting the inner 1/3 of dentin and involving the pulp chamber), YOLOv3 could accurately detect and classify caries with pulpal exposure (class RC6) (IoU₅₀ vs IoU₇₅: precision, 0.77 vs 0.73; recall, 0.61 vs 0.57) but it failed to predict the outer half of enamel caries (class RA1) (IoU₅₀ vs IoU₇₅: precision, 0.35 vs 0.32; recall, 0.23 vs 0.21). Conclusions YOLOv3 yielded acceptable performances in both IoU₅₀ and IoU₇₅. Although the performance metrics decreased in the 7-class detection, the two thresholds revealed comparable results. However, the model could not consistently detect initial-stage caries affecting the outermost surface of the enamel. Clinical relevance YOLOv3 could be implemented to detect and classify dental caries according to the ICCMS™ classification with acceptable performances to assist dentists in making treatment decisions.