Utilizing Artificial Intelligence to Diagnose Temporomandibular Joint Osteoarthritis via Orthopantomographs

2.1. Participants and Sample Size

Radiographic images of the patients with the following selection and exclusion criteria were reviewed.

Selection criteria:

• Visited the Department of Odonto and Stomatology, Hanoi Medical University Hospital with at least 1 of the following clinical symptoms: temporomandibular joint pain, crepitus, and trismus.
• Have panoramic images.
• Accepted to enroll in the study.

Exclusion criteria:

• Had a history of orthopedic surgery.
• Had a history of systemic diseases.

The sample size of the study uses the following formula to evaluate the sensitivity and specificity of a disease diagnosis method:

With:

n_se: Sample size for sensitivity,

: standard normal deviate (α = 0.05 then Z = 1.96)

p_se: Probability of sensitivity. Take p_se = 80% [15].

w: Error of sensitivity and specificity.

p_dis: Prevalence of the disease in the population. Take p_dis = 0.15 [16].

In reality, the research sample size was 651 participants.

Figure 2 illustrates the step-by-step schema of the study.

2.2. Orthopantomogram Image Criteria and TMJOA Diagnosis Criteria

All panoramic radiographs were acquired with standardized imaging parameters: 12 mA tube current, 85 kV tube voltage, and a mean exposure time of 17.6 seconds. Acceptable image quality includes the following:

• Alveolar processes with all upper and lower teeth must be clearly represented.
• The image should show the entire mandible, including the TMJ.
• The dimensions of anatomical structures in vertical and horizontal planes must be symmetrical.
• Density image appearance should be uniform, free of air over the tongue, with the appearance of transparent tape (black) over the roots of upper teeth.
• The panoramic image should not have artifacts due to dentures or removable orthodontic appliances, glasses, earrings, and other jewelry on the patient.

Patients were diagnosed with TMJOA based solely on radiographic evidence of temporomandibular joint changes.

2.3. AI Model Development

An automated model was developed to extract regions of interest (ROI), including the mandibular condyle and surrounding anatomical structures, and to classify them into two categories: normal and osteoarthritis (OA), from each OPG image using a YOLOv11-based object detection framework. Once an image is input into the system, the YOLOv11 model performs inference to detect regions suspected of containing temporomandibular joint abnormalities. The results are presented visually as bounding boxes overlaid on the image, accompanied by predicted class labels and corresponding confidence scores (Fig. 3).

YOLOv11 is a one-stage object detection framework designed for real-time applications by integrating localization and classification into a unified neural network [17]. Unlike two-stage approaches, it eliminates the need for a region proposal mechanism, allowing the model to infer object locations and categories directly from the input image. The detection process begins with preprocessing, where the image is resized and normalized to ensure consistency. During inference, post-processing techniques such as confidence filtering and non-maximum suppression are employed to remove redundant predictions and yield accurate final detections.

Unlike conventional approaches that rely on manual selection of regions of interest (ROIs), the detection-based architecture of YOLOv11 simultaneously localizes the mandibular condyle and performs classification directly on the detected region. This approach is particularly well suited to panoramic radiographs, where the mandibular condyle typically occupies only a small portion of the image and can easily be obscured by surrounding anatomical structures [18, 19]. In this context, whole-image analysis or manual ROI selection may reduce the signal-to-noise ratio and limit the model's ability to focus on diagnostically relevant features. By learning directly from automatically localized regions, YOLOv11 can more effectively capture morphological characteristics associated with TMJOA, including articular surface erosion, subchondral sclerosis, and relative changes in joint space morphology.

The YOLOv11 model was implemented using the PyTorch deep learning framework, enabling efficient model training, optimization, and inference through its dynamic computational graph and GPU acceleration support. All algorithms were programmed and trained on a PC with a GeForce GTX 1080 Ti GPU. Model training was performed using the Adam optimizer with a learning rate of 1.0 × 10⁻⁶. All the training data is divided into mini-batches, with a mini-batch size is set to 16 during training. Training was conducted for up to 1,000 epochs until convergence, with early stopping applied based on the validation loss to prevent overfitting.

Data augmentation techniques, including image rotation (±5 degrees), horizontal and vertical shifts (±10%), brightness adjustment (±10%), and contrast adjustment (±10%), were applied to enhance data diversity and improve model robustness.

2.4. Model and Statistical Analysis

Accuracy, precision, recall, and F1 score were calculated for model performance. Accuracy is defined as the ratio of correct predictions. Precision is the ratio of true positives to true positives and false positives. Recall is the ratio of true positives to true positives and false negatives. Finally, the F1 score is a harmonic mean of precision and recall: (2 × precision × recall)/(precision +  recall). Accuracy, specificity, and sensitivity were calculated for diagnostic performance; Cohen’s kappa was used to estimate agreement in TMJOA diagnoses between experts and AI reads; and McNemar’s test was used to evaluate the significance of the difference. All p values < 0.05 were considered to be statistically significant.

2.5. Ethics Approval and Consent to Participate

The study was approved by the University Council - Hanoi Medical University (Approval No. 4571/QD-DHYHN – 06/10/2023). All procedures involving human participants were performed in accordance with the ethical standards of the institutional and/or research committees and the 1975 Declaration of Helsinki, as revised in 2013. Agreements for participation were requested, and informed consents were given to all participants.

4.1. Comparison with Previous Deep Learning Models

In this study, our deep learning model demonstrated moderate diagnostic performance in the interpretation of panoramic dental radiographs. While slightly lower than some previous benchmarks, these results fall within the range reported in the literature on AI for dental image analysis. One of the earliest deep learning investigations in dental radiography applied a deep neural transfer network (DeNTNet) for the detection of periodontal bone loss in panoramic radiographs [20]. This model, trained on over 12,000 images, achieved an F1 score of 0.75, indicating robust lesion detection and competitive performance relative to clinicians. The variation in performance across studies may be partly explained by design differences. Models like Mask R-CNN, which perform pixel-level segmentation, often provide more precise boundary delineation, enhancing performance on tasks requiring detailed localization compared to object detection-oriented networks that prioritize speed, such as YOLO variants [21]. A systematic review of deep learning in dental radiographs observed that while standard convolutional and segmentation networks generally yield high pooled diagnostic metrics, their efficacy remains highly sensitive to variations in dataset size, imaging modality, and clinical task [22].

4.2. Performance Comparison between AI and Expert

A prominent finding in our study was the marked difference in diagnostic performance between the AI model and expert clinicians. Expert clinicians exhibited high specificity (0.91) but notably low sensitivity (0.51), consistent with well-recognized limitations of OPGs in detecting subtle osseous changes due to anatomical superimposition and low contrast resolution. This discrepancy likely reflects well-documented limitations of OPGs, where overlapping anatomical structures, low resolution, and geometric distortion reduce the visibility of subtle bone changes, making early disease difficult to detect for human observers. These limitations make early condylar bone changes difficult to visualize, even for experienced interpreters, and echo findings from broader literature showing that standard panoramic imaging has limited reliability compared to volumetric modalities such as CBCT for TMJOA evaluation [23]. Previous research confirms that panoramic imaging often underperforms more advanced modalities such as CBCT for bone pathology, including condylar degeneration and other mandibular changes, largely due to these inherent imaging constraints. CBCT provides multiplanar, three-dimensional visualization without superimposition, resulting in superior diagnostic accuracy for changes that may be indiscernible on 2D OPGs [24, 25].

The study’s AI model increased sensitivity from 0.51 of experts to 0.63, identifying true positive cases that were missed by experts. This enhanced ability to detect subtle radiographic patterns mirrored results from previous systematic studies of AI applications in TMJOA detection. A recent meta-analysis of AI models applied to TMJOA radiographic datasets reported pooled sensitivity and specificity values around 0.80 and 0.79, when using both CBCT and panoramic imaging modalities, underscoring that deep learning systems can achieve meaningful diagnostic performance across variable imaging conditions [11]. Furthermore, this review highlighted that architectures such as ResNet exhibit moderate to high sensitivity when trained on adequately annotated condylar osteoarthritis datasets, supporting our finding that AI may overcome some of the human limitations inherent in two-dimensional radiographs. Numerous studies in dental imaging also highlighted the capacity of CNN-based models to extract non-linear, sub-visual features, enabling more sensitive detection of radiographic anomalies such as periapical lesions, bone loss, or other pathologies. In a recent large-scale evaluation of AI performance on panoramic radiographs, deep learning achieved moderate to high sensitivity and specificity for detecting dental conditions, often approaching or exceeding the performance of human readers in certain tasks [26, 27].

Several individual diagnostic accuracy studies further illustrate the performance landscape of AI in TMJOA detection. One study developed deep learning models for condylar osteoarthritis classification using panoramic TMJ projection images and found that models such as GoogLeNet and VGG-16 achieved AUCs up to 0.89, outperforming less experienced human observers, particularly on standardized TMJ projections [28]. Another study focused specifically on OPGs showed that a ResNet-based algorithm yielded sensitivity to 0.73 and specificity to 0.82 when trained and tested in a clinically relevant setting confirmed by CBCT, thus demonstrating performance comparable to experts while achieving a more balanced trade-off [15]. It is also notable that many approaches utilize advanced transfer learning and pre-trained architectures, such as ResNet-152 and EfficientNet-B7, had demonstrated that model selection and training strategy significantly influence diagnostic outcomes [29]. These findings support the notion that AI systems can extract and integrate subtle texture and morphological cues from low-dimensional imaging that may be below the threshold of visual detection by human readers.

Despite these advances, expert clinicians often retain higher specificity, which highlights a key benefit of human judgment in reducing false positives. In our cohort, experts achieved a specificity of 0.91, indicating a strong ability to correctly identify normal temporomandibular joints and avoid overdiagnosis. This finding aligns with multiple radiology and dental imaging studies reporting that experienced clinicians adopt a conservative interpretative threshold, particularly when evaluating panoramic radiographs for subtle degenerative changes. Several factors may explain this pattern.

First, clinicians rely heavily on well-established radiographic hallmarks of TMJOA, such as clear condylar erosion, flattening, osteophyte formation, or sclerosis, and are often reluctant to label early or ambiguous findings as pathological in the absence of unequivocal structural changes [30]. As a result, borderline radiographic variations, age-related remodeling, or projection artifacts are frequently classified as normal by human observers, thereby increasing specificity at the expense of sensitivity.

Second, expert judgment is often influenced by implicit clinical context, even when formal clinical data are not provided. Years of exposure to normal anatomical variability allow clinicians to recognize patterns that are unlikely to represent clinically meaningful TMJOA [31]. This experiential knowledge, while valuable for ruling out disease, may also contribute to underrecognition of early osteoarthritic changes, particularly when structural alterations are mild or atypical [32]. Consequently, expert readers may prioritize diagnostic certainty over early detection, reinforcing high specificity. Several investigations demonstrate that while AI models frequently achieve higher sensitivity, expert clinicians consistently outperform AI in confirming normal joints. This suggests that specificity may represent a strength of human expertise, rooted in holistic interpretation and risk-averse decision-making, whereas AI excels at identifying subtle deviations from learned patterns, regardless of their clinical salience.

From a clinical perspective, this trade-off has meaningful implications. In definitive diagnostic settings or specialist care, high specificity is essential to prevent misclassification and unnecessary escalation of care. However, in screening or primary care, the cost of missed TMJOA diagnoses, particularly in the early stages, may outweigh the drawbacks of increased false positives. In such scenarios, AI-assisted screening could complement expert interpretation by flagging suspicious cases for further review, while clinicians retain ultimate authority over diagnosis. Overall, given the clinical implications of missing an early TMJOA diagnosis, the improved sensitivity of AI models suggests a valuable role for AI as a decision-support tool, particularly in settings where access to CBCT or highly experienced specialists is limited.

Artificial intelligence has emerged as a promising approach in radiology, a field characterized by large volumes of relatively standardized imaging data. Although the current literature remains limited in both quantity and methodological rigor, existing evidence consistently suggests that AI can achieve diagnostic performance comparable to that of healthcare professionals [33–36]. Notably, several studies have reported improved model performance when the training set consisted of ROI images rather than entire radiographs [37]. However, in most published work, ROIs are manually delineated, which reduces the scalability and clinical practicality of AI-based workflows. To address this limitation, the present study implemented an object-detection algorithm to automatically extract ROIs from orthopantomogram (OPG) images.

Previous research using similar techniques has reported exceptionally high precision in mandibular condyle detection, with average precision values of 99.4% on the right side and 100% on the left side [38]. Importantly, the automatically generated ROIs in our study encompassed not only the mandibular condyle but also the articular fossa and articular eminence. This broader anatomical coverage is clinically relevant, as TMJOA is not confined to condylar changes alone. Degenerative alterations are known to involve additional joint components, including the fossa and the eminence, underscoring the need for AI diagnostic models to analyze the entire joint rather than focusing solely on the condyle [39]. Consistent with this rationale, large-scale reviews have reported pooled diagnostic performance with AUC values exceeding 0.90, underscoring the strong clinical potential of AI-based approaches, although variability related to imaging modality and dataset characteristics remains an important factor [13].