Performance of artificial intelligence-based diagnosis and classification of peri-implantitis compared with periodontal surgeon assessment: a pilot study of panoramic radiograph analysis

Abstract

Purpose

The aim of this study was to evaluate the diagnostic and classification performance of a deep learning (DL) model for peri-implantitis–related bone defects using panoramic radiographs, focusing on defect morphology and severity.

Methods

A dataset comprising 1,075 panoramic radiographs from 426 patients with peri-implantitis was analyzed. A total of 2,250 implant sites were annotated and categorized based on defect morphology (intraosseous [class I], supracrestal/horizontal [class II], or combined [class III]) and severity (slight, moderate, or severe). The ensemble-based YOLOv8 DL model was trained on 80% of the dataset, with the remaining 20% reserved for testing. Performance was assessed using classification metrics, including accuracy, precision, recall, and F1 score. The diagnostic accuracy of the DL model was also compared with that of 2 board-certified periodontal surgeons.

Results

The DL model achieved an overall accuracy of 85.33%, significantly outperforming the periodontal surgeons, who exhibited a mean accuracy of 75.6%. The DL model performed especially well for slight class II defects, with precision and recall values of 100% and 98%, respectively. In contrast, the periodontal surgeons demonstrated higher accuracy in severe cases, particularly for class II defects.

Conclusions

DL enables reliable and accurate detection of peri-implantitis bone defects. It outperformed periodontal surgeons in overall accuracy, demonstrating its potential as a valuable second-opinion tool to support clinical decision-making. Future research should focus on expanding datasets and incorporating multimodal imaging.

Graphical Abstract

INTRODUCTION

Peri-implantitis, a significant biological complication associated with dental implants, is characterized by an inflammatory response affecting the alveolar bone surrounding the implant and is often accompanied by progressive bone loss and soft tissue destruction [1]. Recent population-based epidemiological studies have reported wide variations in the frequency of peri-implantitis, with its prevalence ranging from 1.1% to 85.0% and its incidence varying from 0.4% to 43.9% over 5 years [2,3]. The diagnosis of peri-implantitis is primarily based on evaluating the bone level around the implant using dental radiography, along with measuring the peri-implant probing depth. Although these conventional direct and indirect diagnostic methods are widely used and well established, they depend heavily on clinical expertise and are prone to variability due to subjective interpretation and potential inaccuracies [4].

Recent advancements in medical image analysis, particularly the rapid development of deep learning (DL) technology based on convolutional neural networks—a subset of artificial intelligence (AI)—have shown significant promise [5]. Numerous studies have demonstrated that DL techniques can increase diagnostic accuracy, improve workflow efficiency, and support clinical decision-making across various medical applications [6,7]. In dentistry, DL has been employed to analyze both 2- and 3-dimensional dental radiographs, showing considerable potential for improving diagnostic accuracy and clinical outcomes across specialties, including operative dentistry, periodontology, orthodontics, and oral pathology [8,9,10,11].

In the context of dental implants, DL has been used to predict the risk of osseointegration failure and detect fixation fractures [12,13,14]. Furthermore, DL has demonstrated superior performance compared to dental professionals in segmenting and classifying different dental implant systems [15,16]. However, the detection and diagnosis of peri-implantitis around dental implants based on bone morphology and severity remains relatively underexplored in the literature [17]. Thus, the aim of this study was to evaluate the feasibility of using DL techniques to diagnose and classify peri-implantitis–related bone defects on panoramic radiographs, focusing on their morphology and severity.

MATERIALS AND METHODS

Ethics

This study was approved by the Ethics Committee of Jeonbuk National University Hospital (approval No. CUH 2024-11-046) and was conducted in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement and the “AI in dental research” guidelines [18]. As this study utilized a retrospective, de-identified, and anonymized dataset, the requirement for informed consent was waived for all participants.

Dataset

Between January 2018 and October 2024, panoramic dental radiographs of patients with peri-implantitis were collected from the Department of Periodontology at Jeonbuk National University Dental Hospital. Peri-implantitis was diagnosed based on the consensus report of the 2017 World Workshop on the Classification of Periodontal and Peri-Implant Diseases and Conditions [19,20]. From a total of 8,762 panoramic radiographs of patients diagnosed with peri-implantitis, 1,075 radiographs from 426 patients who underwent surgical intervention for peri-implantitis were selected for analysis.

Labeling and preprocessing

From 1,075 panoramic radiographs, a total of 2,250 implant fixtures (excluding prosthetic components) were manually segmented and annotated. The dataset consisted exclusively of implants already diagnosed with peri-implantitis, with their defect morphology and severity pre-identified. To confirm the reliability of the dataset, a single experienced, board-certified periodontal surgeon (Jae-Hong Lee) performed additional validation using the classification criteria outlined in Table 1 [21]. Images were annotated using LabelImg (https://github.com/HumanSignal/labelImg) and Label Studio (https://labelstud.io). Standardization measures were implemented during data preprocessing to ensure consistency and generalizability of radiographic image quality. As imaging conditions vary due to differences in exposure settings, contrast, and resolution, rigorous quality control protocols were applied to minimize potential biases affecting the performance of the DL model. All panoramic radiographs were resized to a standardized dimension of 640×640 pixels to maintain uniform spatial resolution across the dataset, thereby reducing variability due to image acquisition differences. In addition, intensity rescaling was performed by excluding the top 0.2% of pixel values, normalizing contrast and ensuring that variations in exposure settings did not introduce systematic errors. To increase model robustness and mitigate the effects of inconsistent imaging conditions, data augmentation techniques were applied, including random rotations (±20°), horizontal and vertical flipping, and scaling adjustments (±20%).

Table 1. Diagnostic and classification criteria for peri-implantitis, including defect morphology and severity.

Definition
	1.	Presence of bleeding and/or suppuration
	2.	Clinical probing pocket depth of ≥6 mm
	3.	Bone level of ≥3 mm apical to the most coronal part of the implant fixture
Defect morphology
	1.	Class I: Intraosseous defects, including buccal dehiscence, 2- and 3-wall defects, and circumferential defects
	2.	Class II: Supracrestal and/or horizontal defects
	3.	Class III: Combined defects
Defect severity
	1.	Slight: Bone loss of 3–4 mm or <25% of the implant fixture
	2.	Moderate: Bone loss of 4–6 mm or 25%–50% of the implant fixture
	3.	Advanced: Bone loss of >6 mm or >50% of the implant fixture

Model

The choice of the YOLOv8 model for peri-implantitis classification is justified by its superior performance in medical image analysis, particularly in object detection and classification tasks [22]. YOLOv8 offers a balanced combination of speed, accuracy, and computational efficiency, making it well suited for panoramic radiograph analysis that requires precise localization and categorization of peri-implant defects [22,23]. Compared to alternative DL models such as U-Net or Faster R-CNN–based classifiers, the YOLO DL model excels in the real-time processing and detection of complex structures within dental radiographs while maintaining high classification accuracy [24,25]. Its advanced feature extraction capabilities enable effective differentiation of peri-implant defect morphology and severity, which is critical for clinical decision-making. This classification approach is designed to isolate the implant fixture and surrounding alveolar bone, excluding prosthetic components.

For morphological classification, the YOLOv8 classification model categorizes peri-implantitis lesions into 3 classes: intraosseous (class I), supracrestal/horizontal (class II), and combined (class III) defects. This segmentation approach aims to accurately differentiate between the implant fixture-restoration margin and the highest position of the alveolar bone, enabling the calculation of bone resorption percentage. For defect severity, the YOLOv8 segmentation model similarly categorizes peri-implantitis lesions into 3 classes: slight, moderate, and severe defects. Based on the combination of these classes, 9 groups emerged (slight class I/II/III, moderate class I/II/III, and severe class I/II/III). Of the 2,250 implant fixtures, 80% (1,800 fixtures) were assigned to the training and validation dataset, while the remaining 20% (450 fixtures) were assigned to the test dataset. The model was trained for 200 epochs with a batch size of 16. The Adam optimizer was used with an initial learning rate of 0.001, decaying by a factor of 0.1 every 50 epochs. Early stopping was implemented based on the validation loss with patience set at 10 epochs to avoid overfitting (Figure 1).

Figure 1. Architecture of an ensemble-based YOLOv8 deep learning model that integrates backbone and head components to classify the morphology and severity of peri-implantitis defects using panoramic dental radiographs.

SPPF: spatial pyramid pooling fast.

Comparison of the accuracy of DL and periodontal surgeons

The test dataset of 450 implant fixtures, classified into 9 categories of peri-implantitis based on defect morphology and severity, was used to compare the diagnostic accuracy of the DL model with that of 2 board-certified periodontal surgeons. Prior to evaluation by the surgeons, the evaluation dataset was independently reviewed by a researcher who was not involved in the construction (classification, segmentation, and annotation) of the dataset to confirm its reliability.

Statistical analysis

To quantify the classification accuracy of defect morphology and severity in peri-implantitis, various metrics were calculated. Accuracy, defined as (True Positive [TP] + True Negative [TN])/(TP + TN + False Positive [FP] + False Negative [FN]), measures overall model correctness. Precision (TP/[TP + FP]) reflects the proportion of correctly identified positive cases, while recall (TP/[TP + FN]) refers to the model’s capacity to detect TP cases. The F1 score, defined as 2 × ([Precision × Recall]/[Precision + Recall]), provides a balanced measure of model performance by combining precision and recall into a harmonic mean. Receiver operating characteristic (ROC) curves and confusion matrices were used to assess diagnostic performance and visualize classification outcomes. Statistical analyses were performed using Python (version 3.12; Python Software Foundation, Wilmington, DE, USA), the R statistical package (version 4.3.3; R Foundation for Statistical Computing, Vienna, Austria), and SPSS (version 29.0; IBM Corp., Armonk, NY, USA).

RESULTS

Accuracy performance of DL

DL demonstrated an overall accuracy of 85.33%, with precision, recall, and F1 scores consistently ranging from 85.3% to 85.5%. Performance varied according to the severity and morphology of the peri-implantitis defects, with cases of slight severity exhibiting the highest accuracy. Specifically, slight class II defects were associated with the highest diagnostic accuracy (98.0%) and F1 score (99.0%). Moderate and severe defects showed mixed results, with accuracy ranging from 74.0% to 92.0% across these 2 classes. A detailed summary of the DL model’s classification performance is provided in Table 2, while Figure 2 displays the ROC curve and confusion matrix.

Table 2. Comparison of diagnostic accuracy between an ensemble-based deep learning model and periodontal surgeon assessment in the classification of 9 types of peri-implantitis defects.

Defect severity			Defect morphology	Accuracy (%)	Precision (%)	Recall (%)	F1 score (%)
Deep learning
	Total			85.30	85.50	85.30	85.40
		Slight	Class I	90.00	91.80	90.00	90.90
			Class II	98.00	100.00	98.00	99.00
			Class III	72.00	78.30	72.00	75.00
		Moderate	Class I	84.00	85.70	84.00	84.80
			Class II	74.00	84.10	74.00	78.70
			Class III	84.00	85.70	84.00	84.80
		Severe	Class I	92.00	86.80	92.00	89.30
			Class II	92.00	75.40	92.00	82.90
			Class III	82.00	82.00	82.00	82.00
Periodontal surgeon 1
	Total			74.70	81.80	84.60	83.10
		Slight	Class I	68.00	77.50	81.60	79.50
			Class II	64.00	74.40	78.40	76.30
			Class III	62.00	71.80	77.80	74.70
		Moderate	Class I	68.00	71.80	84.80	77.80
			Class II	64.00	73.50	73.50	73.50
			Class III	78.00	84.80	82.40	83.60
		Severe	Class I	92.00	95.70	95.70	95.70
			Class II	94.00	97.90	95.90	96.90
			Class III	82.00	88.60	90.90	89.70
Periodontal surgeon 2
	Total			76.40	83.00	88.80	85.40
		Slight	Class I	72.00	77.80	89.70	83.30
			Class II	78.00	91.70	80.50	85.70
			Class III	82.00	88.60	90.90	89.70
		Moderate	Class I	62.00	71.80	77.80	74.70
			Class II	70.00	70.50	93.90	80.50
			Class III	64.00	66.70	90.90	76.90
		Severe	Class I	88.00	93.50	93.50	93.50
			Class II	92.00	97.90	93.90	95.90
			Class III	80.00	88.40	88.40	88.40

Figure 2. Accuracy performance of deep learning. (A) Receiver operating characteristic curve illustrating the performance of an ensemble-based YOLOv8 deep learning model for the classification of 9 types of peri-implantitis defects. (B) Normalized confusion matrix showing the classification accuracy for each peri-implantitis defect category.

Accuracy performance of periodontal surgeons

The diagnostic accuracy of the periodontal surgeons in classifying the morphology and severity of peri-implantitis defects using panoramic radiographic images was evaluated. Although periodontal surgeon 2 had slightly higher mean scores for accuracy (76.4% vs. 74.7%), precision (83.0% vs. 81.8%), recall (88.8% vs. 84.6%), and F1 score (85.4% vs. 83.1%), the analysis revealed no statistically significant differences in performance metrics between the 2 surgeons. Overall, the surgeons achieved an accuracy of 75.6% across all defect categories, with precision, recall, and F1 scores of 82.4%, 86.7%, and 84.4%, respectively. Accuracy varied by severity and morphology, with 70.0%–72.0% accuracy for slight defects, 65.0%–71.0% accuracy for moderate defects, and up to 93.0% accuracy for severe defects. Notably, severe class II defects demonstrated the highest classification metrics, achieving 93.0% accuracy, 97.9% precision, 94.9% recall, and a 96.4% F1 score (Figure 3).

Figure 3. Accuracy and performance of an ensemble-based YOLOv8 deep learning model compared to periodontal surgeon assessment for the classification of peri-implantitis defects (A-I, slight class I to severe class III): receiver operating characteristic curve analysis of 9 defect types using 450 radiographic images.

DISCUSSION

Peri-implantitis is a complex and serious biological complication that requires accurate diagnosis to prevent implant failure and ensure timely, evidence-based management [26]. A patient-tailored treatment plan combining non-surgical and surgical interventions should consider the defect’s morphology and severity to optimize outcomes and extend implant longevity [27,28]. In the present study, DL achieved a clinically acceptable level of accuracy, surpassing that of the periodontal surgeons by approximately 10%. This finding highlights its potential as a reliable and effective tool in clinical practice.

Two-dimensional dental radiography often underestimates peri-implant bone loss compared to intraoperative measurements and exhibits lower reproducibility for probing depth around implants than natural teeth [29,30]. A cross-sectional study reported significant discrepancies between intraoperatively measured bone levels and radiographically assessed levels (5.3±2.3 mm vs. 4.0±2.2 mm; P=0.014) [29]. Another recent study revealed that intra-examiner reproducibility of probing depth was lower for implants (correlation coefficients, 0.712–0.800) than for natural teeth (0.759–0.863) [30]. Cone-beam computed tomography (CBCT) offers detailed visualization of bone loss patterns; however, adherence to the “as low as reasonably achievable” (ALARA) principle limits the practicality of relying solely on CBCT for diagnosing peri-implantitis [31].

While periapical radiographs offer superior resolution and localized detail, they are limited in assessing overall bone structure and defect morphology at multiple implant sites during a single examination. Panoramic radiographs, despite their slightly lower resolution, provide a wide field of view that allows simultaneous evaluation of multiple implants, facilitating a more comprehensive assessment of peri-implant bone defects. Moreover, using panoramic radiographs reduces radiation exposure compared to the multiple periapical radiographs needed to cover extensive implant sites. Thus, the use of panoramic radiographs is consistent with the ALARA principle, making this approach a practical and patient-centered alternative to higher-dose techniques such as CBCT [32].

Recent DL models have demonstrated high reliability in accurately analyzing various features and patterns on dental radiographs [33]. For instance, YOLOv7 achieved high performance in classifying the severity of peri-implantitis (including bone loss or non-bone loss) on periapical radiographs, with a specificity of 100%, a precision of 100%, a recall of 94.44%, and an F1 score of 97.10% [17]. Another U-Net–based DL detection and segmentation model achieved excellent performance in peri-implantitis analysis, with a segmentation accuracy of 0.999 and classification metrics of 0.777 precision, 0.903 recall, and 0.835 F1 score [34]. Similarly, the ensemble-based YOLOv8 DL model used in this study demonstrated an accuracy of 85.3% in classifying peri-implantitis defects into 9 categories based on morphology and severity.

Compared to previous studies, the present research had the advantage of using a dataset derived exclusively from surgically confirmed cases, increasing its scientific rigor and clinical applicability [17,35]. The comparison between the DL model and the periodontal surgeons in diagnosing the morphology and severity of peri-implantitis defects revealed notable differences in performance metrics. The DL model demonstrated a higher overall accuracy (85.3%) than the surgeons (75.6%), particularly excelling in precision and recall for slight class II defects, where it achieved near-perfect results (100% precision and 98% recall). The superior performance of the DL model in detecting slight class II defects is likely attributable to its capacity to systematically process large datasets and identify subtle radiographic patterns that may be challenging for human evaluators to detect consistently. DL models excel at recognizing well-defined morphological features, especially when structured and homogeneous training data enhance pattern recognition. This high level of accuracy increases diagnostic confidence and reduces inter-observer variability, contributing to more standardized and automated assessments in clinical settings.

Conversely, the periodontal surgeons outperformed the DL model for severe class I and class II defects, with F1 scores of up to 94.6% and 96.4%, respectively. This discrepancy can be attributed to the complexity and heterogeneity of severe defects, which often present with extensive bone loss, irregular radiographic appearance, and overlapping anatomical structures. Unlike AI, which relies primarily on pixel-based learning, periodontal surgeons integrate experiential knowledge, clinical judgment, and multimodal diagnostic inputs—including clinical examination, medical history, and ancillary imaging—to formulate accurate diagnoses [36]. This ability to synthesize diverse clinical information enables human evaluators to better navigate complex and atypical cases, resulting in superior diagnostic performance for severe defect classifications.

Although one surgeon exhibited slightly higher accuracy than the other, the analysis revealed no statistically significant differences between them. This is supported by the Bland-Altman analysis, which indicated an overall mean difference of −0.018 with 95% limits of agreement of −0.083 to 0.073. As these discrepancies fall within an acceptable range, the results suggest that the evaluators demonstrated consistency and maintained inter-rater reliability. The superior performance of the surgeons in severe cases likely reflects the importance of clinical judgment and contextual understanding—an aspect that the DL model may not yet fully replicate. These findings align with prior research and underscore the complementary nature of DL models and human expertise in improving diagnostic accuracy [37,38].

Reliable classification of peri-implantitis defects has key clinical implications for therapeutic decision-making [27,28]. Accurate classification enables the development of tailored treatment plans designed to minimize the risks of under- or over-treatment, including non-surgical, regenerative, or ablative procedures [39]. Early and precise detection of advanced peri-implantitis defects can lead to timely interventions, thus reducing the likelihood of implant failure [39]. By providing a reproducible and objective diagnostic framework, DL may support clinical decision-making while minimizing errors. Furthermore, its reliance on panoramic radiographs is consistent with the ALARA principle; as such, this approach is a practical and patient-centered alternative to techniques like CBCT, which confer a greater radiation dose [32].

The high sensitivity and specificity achieved by the DL model highlight its potential as a valuable second-opinion tool that effectively complements clinicians’ diagnostic skills [40]. AI-based diagnostic tools offer not only enhanced accuracy but also improved clinical efficiency. By automating the labor-intensive process of identifying and classifying peri-implantitis bone defects, clinicians can allocate more time to treatment planning and patient communication. The model can also provide a consistent second opinion, reducing diagnostic errors associated with fatigue or bias. By seamlessly integrating into clinical workflows and providing a consistent framework for decision-making, AI-powered diagnostics have the potential to revolutionize the management of dental implant complications.

The present study has several limitations. First, the dataset was restricted to a narrow range of implant systems collected from 2 dental hospitals, which may not represent the diversity of implant systems and radiographic conditions encountered in broader clinical settings. To address this limitation, further studies should evaluate the feasibility of using DL in real-world practice by incorporating datasets with diverse implant systems and imaging modalities. Second, although selection bias was mitigated by involving an independent investigator for dataset management, the retrospective nature of the dataset could have introduced bias in image selection and annotation. Larger and more heterogeneous datasets are needed to validate these findings and minimize selection bias. Third, reliance on panoramic radiographs, while practical, may limit image sharpness and contrast compared to 2- or 3-dimensional imaging techniques, including periapical radiographs and CBCT. Further research is warranted to explore the utility of datasets based on higher-resolution imaging techniques and to assess the model’s performance across various modalities.

This study demonstrated the potential of AI in diagnosing peri-implantitis by reliably classifying peri-implantitis bone defects based on severity and morphology. The DL model outperformed experienced clinicians in certain diagnostic metrics, particularly for subtle defects. Future studies should validate AI-based diagnostic tools in diverse clinical settings and evaluate their long-term impact on treatment outcomes.