Deep learning versus general dentists: a clinical evaluation of caries diagnostic accuracy on bitewing radiographs

Abstract

Background

The diagnostic accuracy of caries detection on bitewing radiographs varies among dentists and is strongly influenced by lesion severity. Improving the detection of initial and extensive carious lesions remains clinically important for appropriate treatment planning. This study evaluated the diagnostic performance of a deep learning-based detection model and compared its performance with that of general dentists in detecting caries on bitewing radiographs of primary teeth, applying lesion severity criteria relevant to clinical caries management.

Methods

A total of 1,427 bitewing radiographs was included, with 1,180 allocated for training and validation, and 247 for testing. As the reference dataset, two experienced dentists annotated carious lesions according to six depth categories based on the International Caries Classification and Management System (ICCMS™). The diagnostic performance of YOLOv8 and the general dentists were compared using recall, precision, F1-score, average precision (AP), and mean average precision (mAP) at an intersection over union (IoU) threshold of 50%.

Results

YOLOv8 outperformed the general dentists in recall (0.51 vs. 0.29), precision (0.41 vs. 0.31), F1-score (0.44 vs. 0.29), and mAP (0.41 vs. 0.22). Both YOLOv8 and the general dentists demonstrated greater diagnostic accuracy for extensive carious lesions than for initial lesions. Based on the pattern of mispredictions, YOLOv8 tended to underestimate lesion severity, predicting shallower depths than those annotated in the reference dataset, whereas the general dentists exhibited a tendency to overestimate lesion depth.

Conclusion

The deep learning-based model demonstrated superior performance to the general dentists across all evaluated metrics and lesion classes. These findings support the potential role of deep learning as an adjunctive tool for radiographic caries assessment on bitewing radiographs of primary teeth.

Introduction

Detection of proximal caries remains a major clinical challenge during visual examination, particularly in the early stages [1, 2]. Delayed detection may allow lesions to progress into deeper dentin, resulting in extensive tissue destruction and the need for more invasive restorative procedures. Although restorative techniques and materials have advanced considerably in recent years [3, 4], effective caries management still depends on accurate detection to enable timely intervention. Bitewing radiography is an effective method for detecting proximal caries [5–7]. The reported sensitivity of bitewing radiographs for proximal dentin lesions is 0.36 (range: 0.24–0.49), while sensitivity for cavitated proximal lesions exceeds 0.60 [1, 8]. Variability in diagnostic performance may be attributed to differences in clinical experience among examiners and lesion severity [9–12]. To address this challenge, deep learning (DL) has been introduced as a promising adjunctive tool for radiographic caries detection, with growing evidence demonstrating its potential to improve caries detection and classification as well as enhance diagnostic sensitivity among dentists interpreting bitewing radiographs [13–16]. Among these approaches, the DL model “You Only Look Once” (YOLO), a single-step object detection algorithm that localizes targets using bounding boxes, has been explored for caries detection [17–25]. Bounding boxes that contain real objects are considered the ground truth boxes, while those generated by YOLO are referred to as predicted boxes. The accuracy of these predictions is evaluated using the Intersection over Union (IoU) metric, which measures the degree of overlap between the predicted and ground truth boxes. Higher IoU thresholds indicate strong agreement between the predicted and ground truth boxes [20]. Although YOLO models have demonstrated good performance in detecting caries on bitewing radiographs of permanent teeth [19–25], evidence comparing their effectiveness with that of dentists in primary teeth using clinically relevant lesion severity criteria remains limited. While several studies have focused primarily on technical model development, the present study was designed from a clinical perspective to investigate the benefits, limitations, and potential clinical applicability of DL-based caries detection in routine dental practice. Moreover, most studies have evaluated performance using simplified diagnostic tasks, such as caries presence versus (vs.) absence or broad enamel–dentin caries classification, offering limited insight into clinically relevant lesion severity assessment. There remains a need for detection models developed within clinically meaningful frameworks, such as the International Caries Classification and Management System (ICCMS™) [26], which is widely used to guide treatment decisions according to lesion depth. Therefore, this study aimed to evaluate the diagnostic performance of a DL-based model and compare it with that of general dentists in detecting proximal caries on bitewing radiographs of primary teeth using ICCMS™ lesion severity criteria.

Methods

Dataset selection and preparation

This study adhered to the Standards for Reporting of Diagnostic Accuracy Studies (STARD) guidelines [27] and followed the recommended checklist for artificial intelligence (AI) in dental research [28]. The study protocol was approved by the Human Experimentation Committee of the Faculty of Dentistry, Chiang Mai University (approval no. 3/2022), and was conducted in compliance with the Declaration of Helsinki. Following ethical approval and de-identification procedures, bitewing radiographs of primary teeth were retrospectively retrieved from the institutional database of the Faculty of Dentistry, Chiang Mai University. The radiographs had originally been obtained from patients attending the pediatric dental clinic as part of routine dental care for diagnostic evaluation and caries assessment and were not acquired specifically for research purposes. Parents or legal guardians were given the option to provide written informed consent permitting the use of anonymized clinical records and radiographic data for educational and research purposes. Only data from participants for whom such consent had been obtained were included in this study. The study flowchart is illustrated in Fig. 1.

Fig. 1.

Flow of this study

Bitewing radiographs were included if the crowns were clearly visible. The exclusion criteria included permanent teeth, radiographs of poor image quality or incorrect X-ray beam angulation resulting in foreshortening or elongation, proximal surface overlap involving the outer half of enamel or more, and teeth presenting enamel or dentin abnormalities such as amelogenesis imperfecta, dentinogenesis imperfecta, or dentin dysplasia. The radiographs were acquired using a dental X-ray unit (Heliodent Plus™, Dentsply Sirona Corp., USA, 70 kV, 10 mA, exposure time 0.1 s) with Carestream 7600 photostimulable phosphor plates (Carestream, Rochester, NY, USA). The number of radiographs for the test dataset was calculated by sample size estimation from https://www.openepi.com/DiagnosticTest. The estimated sensitivity was based on outcomes of a previous study that assessed caries classification in primary teeth through the use of a DL-based model [29]. Consequently, the test dataset consisted of a minimum of 433 tooth surfaces. To account for clustering, we calculated the Design Effect (DE) using the formula DE = 1 + (m-1) * ICC. The intraclass correlation coefficient (ICC) was assumed to be 0.57, as previously reported in the literature [30]. “m” denotes the cluster size, which we estimated at 10 proximal surfaces per bitewing radiograph. As a result, the total number of test datasets comprised at least 2,468 tooth surfaces, equivalent to 247 bitewing radiographs.

From a total of 4,269 eligible bitewing radiographs acquired between 2017 and 2022 from the Synapse Radiology Picture Archiving and Communication System (version 5.7.100; FUJIFILM Medical Systems, U.S.A., Inc.), 1,427 radiographs were selected—comprising 247 for the independent test dataset and 1,180 for the training/validation dataset—using computer-generated random sampling stratified by year without including any patient-identifying information, yielding approximately 237–238 radiographs per year.

Reference dataset

All proximal lesions in the primary teeth were annotated by two experienced dentists—an oral radiologist (W.P.) and a pediatric dentist (O.W.) with 10 and 21 years of clinical experience, respectively—using LabelImg (LabelImg, 2015; http://github.com/tzutalin/labelImg) on a medical-grade liquid crystal display (LCD) monitor in a dimmed-light radiology workstation [31, 32]. Before study initiation, both experienced dentists underwent a calibration process involving protocol alignment and standardized application of the caries classification criteria to ensure consistent interpretation and annotation. Intra- and inter-examiner reliability were evaluated using ICCs. The intra-examiner reliability values were 0.935 (W.P.) and 0.928 (O.W.), while the inter-examiner reliability was 0.913, indicating excellent agreement. Each carious lesion was identified by enclosing it within a rectangular ground-truth bounding box fitted to the lesion size, and lesion depth was classified into six levels according to the ICCMS™ [26], as shown in Table 1. These annotations were used as the gold standard reference dataset for evaluating the diagnostic performance. To minimize the risk of incorporation bias, the reference standard was established independently through a consensus-based annotation process. Neither the outputs of the DL model nor the interpretations of the general dentists were incorporated into the construction of the reference standard. In cases of disagreement, a third experienced oral radiologist independently reviewed the lesions to establish consensus. Only six lesions required third-examiner adjudication, resulting in an overall agreement rate of 99.90%.

Table 1.

Radiographic scoring criteria based on the International Caries Classification and Management System (ICCMS™) [26]

Model training and validation

YOLOv8 was selected as a representative model due to its high performance and wide range of capabilities, including detection, classification, segmentation, pose estimation, and tracking [33]. The training process for YOLOv8 was conducted using a dataset comprising 1,180 labeled bitewing radiographs. The model was trained using a 4-fold cross-validation approach. For each fold, the radiographs were divided into training and validation sets (845/294, 849/291, 841/293, and 839/290, respectively) to maintain a balanced distribution of carious lesions across all depth classes, ensuring that the model was uniformly exposed to each class and that the validation and test results reflected realistic performance across classes. YOLOv8 was trained using an input size of 1280 × 1280 × 3 for 400 epochs with a batch size of 4, employing the AdamW optimizer (learning rate = 0.001, momentum = 0.937, weight decay = 0.0005), early stopping with a patience of 50 epochs, and evaluation thresholds set at an IoU of 0.50 and a confidence level of 0.25. Regarding the handling of class imbalance during training, focal loss and data augmentation techniques were implemented. The YOLOv8 model utilized focal loss as the primary loss function, which is specifically designed to down-weight frequent examples (majority classes) and focus the learning process on less frequent cases (minority classes). In addition, data augmentation was applied during training to enhance the model’s exposure to minority classes. The first fold was selected for the testing process due to its highest performance.

Human comparators

Human comparator recruitment commenced with an eligibility screening that required experience in caries detection on bitewing radiographs as part of routine clinical practice, and successful completion of a qualification. The qualifying examination comprised 30 caries lesions (five per class based on the ICCMS™ system), with a minimum passing score of 60%, which was set by the researchers according to the findings of Panyarak et al., who studied the performance of YOLOv3 and YOLOv7 on caries detection in bitewing radiographs in permanent teeth based on the ICCMS™ radiographic scoring and reported a recall of 55.2 (YOLOv3) vs. 69.7 (YOLOv7) [21]. Five general dentists meeting these criteria were enrolled as human comparators. The participating dentists were general practitioners with 2 to 6 years of clinical experience who had received training in the ICCMS™ radiographic scoring during undergraduate education and routinely used it for evaluating carious lesions on bitewing radiographs. Before the testing process, all participants underwent a standardized training program comprising a theoretical session on ICCMS™ radiographic scoring, a calibration exercise using exemplar cases, and a written handbook outlining the testing procedures and annotation protocol.

Testing process

A total of 247 bitewing radiographs were used as the test dataset to evaluate the performance of YOLOv8 and the general dentists. Each dentist independently annotated a subset of 48–50 radiographs under the same protocol and viewing conditions used to establish the reference dataset. Prior to the testing, the distribution of lesion classes across the subsets assigned to each dentist was statistically evaluated using Pearson’s chi-square test. The results showed no significant differences among the subsets (χ² = 18.11, df = 20, p = 0.58), indicating a comparable and balanced distribution of lesion classes across examiners. Intra-examiner reliability was assessed using 10% of the test dataset, which was re-evaluated by each examiner one month after the testing process.

Evaluation metrics and statistical analysis

To evaluate object detection performance, true positives (TP) were defined as predicted bounding boxes that matched the ground truth, with correct classification and IoU ≥ 50%. False positives (FP) included predictions with IoU < 50%, correct localization but incorrect classification, or erroneous detections involving non-lesion structures such as sound tooth surfaces, bite tabs, or soft tissue. False negatives (FN) referred to lesions which were not detected by the model or dentists. True negatives (TN)—correct identifications of non-object regions—are not reliably measurable in object detection due to the impracticality of enumerating all potential negative instances within an image; therefore, metrics such as accuracy and specificity are not applicable in this context [34]. Performance was assessed using standard object detection metrics, including recall, precision, F1-score, average precision (AP), and mean average precision (mAP) [20, 21].

Recall, or Sensitivity, reflects the ability to correctly detect and classify carious lesions relative to the total number of lesions at each caries depth level. Higher recall means fewer missed lesions; a recall of 1.0 means every true caries was detected. It was calculated as:

Precision, or Positive Predictive Value (PPV), represents the proportion of correctly identified positives among all positive predictions and reflects the likelihood that a predicted carious surface truly has caries. Higher precision means fewer FP; a precision of 1.0 means every detected lesion was correct. It was calculated as follows:

F1-score provides a harmonic balance between precision and recall. High value indicates the evaluator is both accurate in detection and reliable in avoiding false alarms. It was calculated as:

Average Precision (AP) represents the area under the interpolated Precision-Recall (PR) curve, with higher values indicating strong performance—characterized by both high precision and recall—across varying confidence thresholds. The modified AP calculation used in this study, as reported by Panyarak et al. [20], is as follows:

Pinterp(r) is the highest precision value at each recall level from the interpolated PR curve, r represents recall and p(ṙ) represents precision at recall level ṙ. This value was calculated as:

Pinterp (r) = Maxp (r ̇)

Mean average precision (mAP) is the average of AP values from all classes. The formula is as follow [20]:

APi is interpolated AP of object category.

The diagnostic results from all dentist-evaluated subsets were pooled using a micro-averaging approach. All TP, FP, and FN were aggregated across all evaluators to represent the collective performance of the general dentist group, rather than averaging the individual performance scores of each dentist. The performance metrics of the general dentist group for each class were calculated as follows, with recall shown as an example. Precision, F1-score, and AP were calculated using the same approach.

To reduce the influence of class imbalance, a macro-averaging approach was used for calculating overall performance metrics. Performance for each lesion class was calculated independently and then averaged with equal weighting across classes. This approach was intended to minimize disproportionate influence from more prevalent lesion classes, such as RA3, and to provide a more balanced assessment of performance across all lesion categories. The overall performance metrics were calculated as follows, with recall shown as an example. Precision, F1-score, and mAP were calculated using the same approach.

Bootstrap-derived 95% confidence intervals (CIs) were calculated for the diagnostic performance metrics of YOLOv8 and the general dentists to enable statistical comparison [35]. Confidence intervals were estimated using the bias-corrected and accelerated bootstrap method with 1,000 iterations for CI. Differences in performance between YOLOv8 and the aggregated dentist group were interpreted based on overlap of the corresponding CIs. Intra-examiner reliability for all examiners and the experienced dentists, as well as inter-examiner reliability between the experienced dentists, was evaluated using the ICC with a two-way mixed-effects model (single measures, absolute agreement). Statistical analyses were performed using Statistical Package for the Social Sciences (SPSS) software, version 19.0 (IBM Corp., Armonk, NY, USA).

Results

A total of 1,427 bitewing radiographs were allocated into a training/validation dataset (1,180 films) and a testing dataset (247 films), with both datasets showing the highest prevalence of RA3-level carious lesions. The detailed lesion distribution is provided in Table 2.

Table 2.

Distribution of carious lesions in the training/validation and testing datasets across six classification categories

Classification		Training/Validation dataset	Testing dataset
		N (%)	N (%)
Initial stage: RA	RA1	636 (12.39)	175 (18.50)
	RA2	1,183 (23.05)	161 (17.02)
	RA3	1,437 (28.00)	212 (22.41)
Moderate stage: RB	RB4	978 (19.05)	163 (17.23)
Extensive stage: RC	RC5	481 (9.37)	120 (12.68)
	RC6	418 (8.14)	115 (12.16)
All classes		5,133 (100)	946 (100)

A comparative analysis of caries detection across all lesion depths revealed that YOLOv8 consistently outperformed general dentists across all evaluated metrics, as illustrated in Fig. 2. Detailed analysis by lesion class (RA1–RC6) revealed that YOLOv8 consistently outperformed the general dentists across all evaluation metrics, with particularly pronounced gains in recall and AP (Fig. 3). YOLOv8 and the general dentists achieved higher performance in detecting and classifying more advanced lesions. For initial-stage lesions (RA), YOLOv8 demonstrated notably higher recall (RA1: 0.24 vs. 0.13; RA2: 0.45 vs. 0.15; RA3: 0.63 vs. 0.28), F1-score (RA2: 0.41 vs. 0.18), and AP (RA2: 0.35 vs. 0.15; RA3: 0.42 vs. 0.18). In the moderate stage (RB), YOLOv8 maintained superior recall (RB4: 0.60 vs. 0.41) and AP (RB4: 0.41 vs. 0.27), with F1-scores that were marginally higher or comparable to those of dentists. For extensive-stage lesions (RC), YOLOv8 significantly outperformed dentists across all metrics, with RC6 exhibiting the greatest precision advantage (0.62 vs. 0.35) and a substantial gain in AP (0.61 vs. 0.27). Overall, precision values for both comparators were generally low and comparable across most lesion depths, except for RC6, where YOLOv8 achieved its most pronounced advantage, as additionally evident in the precision–recall curve shown in Fig. 4. These findings indicate that YOLOv8 not only identified a greater number of true lesions but also maintained competitive precision, particularly in the detection of advanced lesions.

Fig. 2.

Overall caries detection performance of YOLOv8 and general dentists at an intersection over union of 50%. YOLOv8 demonstrated superior performance across all evaluated metrics, including recall (0.51 vs. 0.29), precision (0.41 vs. 0.31), F1-score (0.44 vs. 0.29), and mean average precision (mAP) (0.41 vs. 0.22), compared with the general dentists

Fig. 3.

Caries detection performance of YOLOv8 and general dentists in six-class classification at an intersection over union of 50%. YOLOv8 demonstrated superior performance across most lesion classes, particularly for initial lesions, with higher recall for RA1 (0.24 vs. 0.13), RA2 (0.45 vs. 0.15), and RA3 (0.63 vs. 0.28). YOLOv8 also achieved higher average precision (AP) for initial and moderate lesions (RA2: 0.35 vs. 0.15; RB4: 0.41 vs. 0.27) and higher precision and AP for extensive lesions (RC6) (0.62 vs. 0.35 and 0.61 vs. 0.27, respectively)

Fig. 4.

Precision-recall curves illustrating the diagnostic performance of YOLOv8 across lesion depths (RA1–RC6) at an intersection over union ≥ 50%. YOLOv8 showed improved performance with increasing lesion depth, achieving the highest average precision in RC6 (0.608). The overall mean average precision across all classes was 0.414

The percentage-normalized confusion matrices demonstrate the predictions of YOLOv8 and the general dentists in each class of caries depth in detail (Fig. 5). In general, both groups demonstrated good alignment with the reference dataset, with YOLOv8 outperforming the dentists in terms of true positives across all classes, except for RC6. Based on the pattern of mispredictions, YOLOv8 tended to underestimate lesion severity, predicting shallower depths than those annotated in the reference dataset, whereas the general dentists exhibited a tendency to overestimate lesion depth. The results also showed that YOLOv8 generated more false-positive detections by misidentifying background as carious lesions across all classes, as illustrated in the rightmost column of the confusion matrix. When examining undetected caries lesions, indicated by the bottom background row of the confusion matrix, YOLOv8 exhibited fewer false negatives across all caries depths. Intra-examiner reliability, reassessed after a one-month interval, is presented in Table 3.

Fig. 5.

Percentage-normalized confusion matrices of YOLOv8 and general dentists for six-class ICCMS™ caries classification. ‘True’ represents annotations from the reference dataset, and ‘Predicted’ denotes the outcomes from each examiner. The ‘background’ in the predicted rows indicates undetected carious lesions, whereas the ‘background’ in the true columns represents non-carious areas, including sound tooth structures and non-tooth regions. Darker cells indicate a higher number of predictions, with diagonal cells representing correct classifications. YOLOv8 showed higher correct classification rates for several lesion classes, particularly RA3 (57.08% vs. 28.30%), RB4 (47.85% vs. 36.81%), and RC5 (41.67% vs. 32.50%), compared with the general dentists

Table 3.

Intra-examiner reproducibility of diagnostic performance among five general dentists

	General dentist 1	General dentist 2	General dentist 3	General dentist 4	General dentist 5
ICC [95% CI]	0.94 [0.86, 0.98]	0.90 [0.76, 0.96]	0.86 [0.76, 0.95]	0.87 [0.71, 0.98]	0.92 [0.81, 0.97]

ICC intraclass correlation coefficient, CI confidence interval

Over-detections or FP among examiners were attributed to the unique anatomy of primary teeth, including the mesial radiolucency of maxillary first primary molars resulting from concavities, as well as cervical burnout, and the Mach band effect. YOLOv8 also made errors, including false-positive predictions on non-tooth structures and difficulty detecting multiple lesions per tooth (Fig. 6).

Fig. 6.

The examples of incorrect detection by YOLOv8: (a) YOLOv8 incorrectly detected radiolucency in bite tabs; (b) YOLOv8 incorrectly detected radiolucency in soft tissue; (c) YOLOv8 struggled to distinguish multiple carious lesions in a single tooth

The mean performance estimates and their corresponding bootstrap-derived 95% CIs are presented in Table 4 and graphically illustrated in Fig. 7. These intervals were used to assess statistical significance between YOLOv8 and general dentists. YOLOv8 demonstrated higher performance than the aggregated general dentist group across all evaluated metrics, achieving recall, precision, F1-score, and mAP values of 0.51 (95% CI: 0.49–0.53), 0.39 (95% CI: 0.36–0.41), 0.44 (95% CI: 0.43–0.45), and 0.35 (95% CI: 0.33–0.37), respectively. The corresponding values for the general dentists were 0.31 (95% CI: 0.29–0.32), 0.31 (95% CI: 0.29–0.32), 0.30 (95% CI: 0.29–0.31), and 0.13 (95% CI: 0.11–0.14). Non-overlapping CIs were observed for all metrics.

Table 4.

Diagnostic performance comparison between YOLOv8 and the general dentists on the test dataset, presented as mean estimates with corresponding bootstrap-derived 95% confidence intervals

Comparators	Mean [95% CIs]
	Recall	Precision	F1-score	mAP
YOLOv8	0.51 [0.49, 0.53]	0.39 [0.36, 0.41]	0.44 [0.43, 0.45]	0.35 [0.33, 0.37]
General Dentists	0.31 [0.29, 0.32]	0.31 [0.29, 0.32]	0.30 [0.29, 0.31]	0.13 [0.11, 0.14]

CI confidence interval, mAP mean average precision

Fig. 7.

Bootstrap-derived 95% confidence intervals for recall, precision, F1-score, and mean average precision (mAP) for YOLOv8 and the general dentists. YOLOv8 achieved higher recall (0.51 vs. 0.31), precision (0.39 vs. 0.31), F1-score (0.44 vs. 0.30), and mAP (0.35 vs. 0.13) than the general dentists

Discussion

Deep learning has demonstrated considerable promise in caries detection, to the extent that it may serve as a valuable adjunctive aid in clinical practice. Our findings demonstrated that YOLOv8 outperformed general dentists across all evaluation metrics, consistent with previous comparative studies. Frutos et al. [22] reported that YOLOv5 outperformed the best annotators (dental clinicians), with superior mAP (0.647 vs. 0.299), F1-score (0.548 vs. 0.495), and a lower FN rate (0.149 vs. 0.164). Likewise, Ayhan et al. [23] demonstrated that YOLOv9c outperformed six dentists, with higher recall (0.727 vs. 0.269–0.613) and F1-score (0.687 vs. 0.394–0.685). These studies, along with our results, underscore the potential of well-trained YOLO to detect caries as effectively as, or even better than, human comparators across varying levels of clinical experience, task complexity, and model architectures.

Greater task complexity can lead to lower performance, even when using the same evaluation metrics. Several studies reporting high diagnostic performance assessed only binary caries presence or absence, overlooking the complexity of real-world diagnostic tasks routinely undertaken by dentists. This study was therefore designed with greater diagnostic difficulty, requiring the model to detect and classify lesions across six ICCMS™ stages. To broaden applicability, we focused on primary teeth, whose anatomical variations often resemble caries on radiographs, presenting challenges even for experienced clinicians. Consequently, the evaluation metrics observed in this study were relatively modest compared with previous studies [19–25].

YOLOv8 demonstrated superior recall in detecting initial-stage lesions (RA), indicating greater sensitivity in identifying lesions that are often difficult for the human eye to detect. A recent study by Bayati et al. demonstrated the effectiveness of YOLOv8 in detecting early-stage carious lesions on bitewing radiographs, achieving high recall and precision [25]. Previous studies have similarly reported strong performance of DL models compared with dentists. YOLOv9c outperformed dentists in recall for enamel caries detection and achieved an F1-score comparable to the best-performing experienced dentist [23]. Likewise, U-Net demonstrated robust sensitivity (0.70) for enamel caries, whereas most dentists showed low sensitivity, with all but one recording values below 0.25 [13]. Early caries detection is critical, as it enables the timely implementation of preventive and non-restorative strategies, including remineralization and active monitoring at appropriate recall intervals.

Although RC6 comprised only 8.14% of the training set, YOLOv8 achieved high recall for RC6 lesions. This finding is consistent with the performance of the general dentists and previous studies [22, 23], suggesting that severe lesions are inherently easier to recognize because of their more distinct radiographic features. Interestingly, the model’s strong performance on RA3 lesions may be explained by their higher prevalence in the dataset, underscoring the influence of data distribution on model learning. Comparable trends were noted by Frutos et al. [22], who reported higher mAP and F1-scores with YOLOv5 when trained on a much larger dataset of 13,887 bitewings, highlighting how both lesion severity and dataset size shape diagnostic performance. Despite potential effects of class prevalence on AP and mAP, imbalance was mitigated through focal loss, data augmentation, and macro-averaging. The high AP for the underrepresented RC6 class suggests that model performance reflected lesion characteristics rather than class frequency. The bootstrap analysis strengthened the robustness of the performance comparison by accounting for sampling variability within the test dataset. Although 1,000 resampling iterations resulted in changes in some point estimates, such as mAP, these differences reflect performance under repeated sampling and provide a more robust assessment of model stability and uncertainty. Notably, the consistently non-overlapping 95% CIs across all metrics further support the superior and stable diagnostic performance of YOLOv8 relative to the general dentist group. The confusion matrices revealed differing prediction tendencies, with YOLOv8 prone to underestimation and general dentists to overestimation. A similar tendency was reported by Giannakopoulos et al. (2026), who found that the AI software Diagnocat™ was more reliable for excluding disease than for confidently confirming lesion presence, regardless of lesion depth [36]. This complementarity suggests that using AI as an adjunct, while reserving final clinical judgment for dentists, may provide the greatest benefit.

Model architecture is recognized as an important factor influencing caries detection performance. Panyarak et al. [21], applying the ICCMS™ system on permanent teeth, reported that YOLOv7 outperformed YOLOv3 in precision, F1-score, and mAP, though with lower recall. More recently, Alsolamy et al. (2025) reported superior detection performance of YOLOv11 across initial (RA), moderate (RB), and extensive (RC) caries stages [24] compared with results previously reported for YOLOv3 within a similar four-class ICCMS™ classification framework [20]. These findings further support the influence of architectural advances on AI-assisted caries detection. Nevertheless, newer model architectures do not invariably guarantee improved outcomes. Previous comparative work has shown YOLOv8 to achieve the highest detection accuracy (mAP at IoU 50% = 0.982) over YOLOv7 (0.721) and YOLOv9 (0.832) [37]. YOLOv8 was selected in this study as a representative model because it was the state-of-the-art YOLO architecture at the time the study was initiated. This study focused on evaluating a contemporary model within a clinically relevant diagnostic framework rather than on model optimization. Given the rapid evolution of DL technologies, newer YOLO-based models with continuously advancing capabilities may further enhance caries detection performance.

The false-positive detections observed in this study, particularly in non-tooth structures and radiolucent anatomical regions, represent a limitation for clinical implementation. In clinical workflow settings, excessive false-positive findings may increase the need for clinician verification, reducing efficiency and potentially causing unnecessary concern. In addition, frequent false-positive detections may affect clinicians’ confidence in the reliability of the model and could influence their willingness to adopt AI-assisted systems in routine clinical practice. Future model refinement may benefit from post-processing strategies, such as tooth segmentation or anatomical filtering, to restrict lesion detection to clinically relevant regions and reduce false-positive rates. Incorporating hard-negative examples, secondary verification models, or multi-stage frameworks may further improve diagnostic specificity and clinical applicability.

An area for improvement in this study was the limited training dataset and lack of histological validation; however, these were addressed through a fuzzy gold standard based on annotations from the experienced dentists. As this study was retrospective and based on a dataset of 1,427 bitewing radiographs, histological confirmation or tooth separation was not feasible. This limitation appears to be widespread in the current literature. A recent systematic review of DL approaches for radiographic caries detection on bitewing radiographs reported that the reference standard domain consistently exhibited a high risk of bias, as all included studies published between 2020 and 2025 relied on expert annotations without histological validation [16]. Although a recent study using histological evaluation of 548 proximal surfaces as the reference standard demonstrated AI performance comparable to human raters and superior sensitivity for early enamel lesions, the findings were derived from an ex vivo setting [38]. Consequently, differences between ex vivo and clinical imaging conditions may limit direct comparability and introduce domain-shift effects, highlighting the continuing need for clinically validated datasets and external validation studies. The superior performance observed in the present study should be interpreted within the context of radiographic assessment. Because the reference standard was based on the evaluation of bitewing radiographs, the findings reflect performance in radiographic lesion detection and classification rather than comprehensive clinical diagnosis. Therefore, the proposed model should be regarded as an adjunctive tool for radiographic interpretation and decision support rather than a replacement for clinical examination and professional judgment.

The ICCMS™ was selected because it is an internationally recognized, evidence-based system that is widely applicable in clinical practice and provides clearly defined criteria for lesion depth classification, thereby reducing subjectivity and potential bias. The reproducibility of this scoring system has been reported to be substantial or excellent. The accuracy when related to histology using Spearman correlation coefficients was excellent [26, 39]. However, using the same conceptual framework (ICCMS™) across the reference experienced dentists, the model, and the human comparators may introduce the possibility of shared systematic bias. To minimize potential bias, several methodological measures were implemented. The general dentists and the model were evaluated independently against the reference dataset, and neither their interpretations nor the model outputs contributed to the construction of the reference standard. Moreover, the experienced dentists underwent independent calibration, and all examiners received standardized training to ensure consistent application of the ICCMS™ criteria, thereby reducing reliance on individual subjective thresholds. Strict separation between training and testing datasets was maintained, and imaging as well as reading conditions were standardized. However, the use of a shared imaging modality and diagnostic framework may introduce residual systematic bias. Furthermore, this study compared YOLOv8 with the aggregated performance of general dentists to better reflect the range of diagnostic variability encountered in routine clinical practice rather than relying on a single examiner. Although examiner variability was minimized through qualification procedures, standardized ICCMS™ training, and balanced lesion-class distributions across subsets, the use of different test subsets for each dentist may still limit direct examiner-level comparability. The single-center design and use of a single imaging system may also limit external validity and introduce potential device-specific performance bias, as DL models can be sensitive to variations in image acquisition, processing algorithms, and detector characteristics. These limitations should be considered when interpreting the findings. Limiting the analysis to primary dentition may restrict the generalizability of the findings. Further studies involving mixed and permanent dentitions, as well as external multi-center datasets, are needed to validate the robustness and broader applicability of the proposed model.

Future work should incorporate larger and more diverse training datasets, clinical validation methods such as tooth separation, and multi-algorithm frameworks with broader optimization strategies. Clinically oriented applications, including real-time detection within artificial intelligence–Internet of Things (AI–IoT)-enabled dental workflows, also warrant further investigation. Despite these limitations, the findings suggest that DL-based detection has the potential to serve as an adjunctive tool for radiographic interpretation, improving the consistency of caries lesion assessment, particularly for early lesions that remain challenging to detect [13].

Conclusions

Within the scope of radiographic caries assessment, the DL-based model demonstrated consistently higher performance than general dentists across all evaluated metrics and lesion classes. These findings support the potential role of DL as an adjunctive tool for radiographic interpretation of caries lesions and lesion severity assessment on bitewing radiographs of primary teeth. However, the model should be regarded as a complement to, rather than a substitute for, comprehensive clinical examination and professional judgment.

Abbreviations

ICCMS™

International Caries Classification and Management System

AP

Average precision

mAP

Mean average precision

IoU

Intersection over union

DL

Deep learning

YOLO

You Only Look Once

STARD

Standards for Reporting of Diagnostic Accuracy Studies

AI

Artificial intelligence

DE

Design effect

ICC

Intraclass correlation coefficient

LCD

liquid crystal display

TP

True positives

FP

False positives

FN

False negatives

TN

True negatives

PPV

Positive predictive value

PR curve

Precision-recall curve

CI

Confidence Interval

Authors’ contributions

Conceptualization and Methodology: P.K., W.P., W.S., A.N., K.W., O.W.; Data collection and data preparation: P.K., O.W., W.P.; Data annotation for reference dataset: O.W., W.P.; Programming, implementation of the computer code, model training and testing: W.S.; Data analysis: P.K., W.P., W.S., A.N., K.W., O.W.; Statistical analysis: P.K., A.N., W.S., O.W.; Writing – original draft preparation: P.K., O.W.; Writing – review and editing: P.K., W.P., W.S., A.N., K.W., O.W.; Supervision and Administration: O.W.; All authors read and approved the final manuscript.

Funding

This study was supported by a postgraduate research scholarship from the Faculty of Dentistry, Chiang Mai University.

Data availability

The datasets generated and/or analyzed during the current study are not publicly available because they contain patient-related information and are subject to ethical and institutional restrictions. De-identified data supporting the findings of this study may be available from the corresponding author, Onnida Wattanarat (onnida.wattanarat@cmu.ac.th), upon reasonable request and with approval from the Human Experimentation Committee, Faculty of Dentistry, Chiang Mai University. The code and trained models used in this study are available from Wattanapong Suttapak (wattanapong.su@up.ac.th) upon reasonable request.

Declarations

Ethics approval and consent to participate

This study was approved by the Human Experimentation Committee of the Faculty of Dentistry, Chiang Mai University (approval no. 3/2022), and was conducted in compliance with the Declaration of Helsinki. Written informed consent was obtained from the parents or legal guardians of all participants. The study design adhered to the recommended checklist for artificial intelligence in dental research and the Standards for Reporting of Diagnostic Accuracy Studies (STARD) guidelines.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.