A Deep Learning Approach for Mandibular Condyle Segmentation on Ultrasonography

Abstract

Deep learning techniques have demonstrated potential in various fields, including segmentation, and have recently been applied to medical image processing. This study aims to develop and evaluate computer-based diagnostic software designed to assess the segmentation of the mandibular condyle in ultrasound images. A total of 668 retrospective ultrasound images of anonymous adult mandibular condyles were analyzed. The CranioCatch labeling program (CranioCatch, Eskişehir, Turkey) was utilized to annotate the mandibular condyle using a polygonal labeling method. These annotations were subsequently reviewed and validated by experts in oral and maxillofacial radiology. In this study, all test images were detected and segmented using the YOLOv8 deep learning artificial intelligence (AI) model. When evaluating the model’s performance in image estimation, it achieved an F1 score of 0.93, a sensitivity of 0.90, and a precision of 0.96. The automatic segmentation of the mandibular condyle from ultrasound images presents a promising application of artificial intelligence. This approach can help surgeons, radiologists, and other specialists save time in the diagnostic process.

Introductıon

Artificial intelligence (AI) refers to the use of computers and machines to perform tasks that typically require human intelligence [1–5]. Through machine learning and deep learning—two key subsets of AI—computers and robots can be trained to analyze specific types of data using various techniques. AI is now widely applied across multiple industries, including manufacturing, financial markets, and healthcare. AI-driven programs are capable of processing vast amounts of data collected from different sources [3–6]. However, despite its growing presence, the full scope and impact of AI—both personally and professionally—remain largely unknown to many individuals, including scientists and medical professionals.

In dentistry, AI has seen remarkable and expanding applications, particularly in the field of radiology [2–9]. Recent studies have explored its potential in various areas, such as diagnosing dental caries, detecting diseases, planning orthodontic treatments for misaligned teeth, and employing robotic surgery for dental implant placement [10–14]. AI’s ability to work with image processing techniques has significantly advanced dental radiology research. Current studies focus on applications such as classifying and segmenting teeth in two-dimensional (2D) and three-dimensional (3D) radiographic images, identifying dental and periodontal diseases, evaluating risk groups, automatically marking anatomical structures, conducting cephalometric analyses, and diagnosing conditions like osteoporosis, which can be detected in radiographs [5]. Additionally, AI has been utilized for diagnosing and planning treatments for various orofacial disorders, as well as for early screening of oral cancer and cervical lymph node metastases [7, 8].

The temporomandibular joint (TMJ) consists of several anatomical components, including the mandibular condyle, temporal bone, articular disc, ligaments, and associated muscles. It is a ginglymoarthrodial joint, capable of both hinge and sliding movements. Positioned beneath the temporal bone, it is located between the mandibular fossa and the mandibular condyle, in front of the external auditory canal [15–19]. To ensure smooth articulation, a biconcave-shaped articular disc, composed of dense fibrous connective tissue, sits between the mandibular fossa of the temporal bone and the condylar head of the mandible [16, 19, 20]. The joint is lubricated by synovial fluid, which contains mucopolysaccharides that provide essential metabolic support to the tissues. Since the articular surfaces and disc lack blood vessels, the synovial fluid plays a critical role in nourishing the joint structures [18, 19, 21–23]. The mandibular condyle, a vital part of the TMJ, regulates the mandible’s sagittal and vertical growth and is essential for speaking, swallowing, and mastication. The condyle’s shape and integrity are clinically important, as various conditions can alter it—for example, osteoarthritis can cause surface erosion or flattening, trauma can lead to fractures or deformities, and developmental anomalies like condylar hyperplasia or bifid condyle affect its form. Accurate delineation of the condyle is therefore valuable for diagnosing TMJ disorders and planning treatments. In TMJ osteoarthritis, precise segmentation of the condyle allows quantitative assessment of bone changes (e.g., volume, surface irregularities), aiding in evaluating disease severity [24]. Similarly, for surgical interventions, such as TMJ arthroscopy or arthrocentesis, identifying the condyle’s exact location is essential for instrument placement. Ultrasound (US) guidance during TMJ injections has been shown to improve accuracy, reduce needle placement attempts, and minimize trauma—automated condyle segmentation on US could further enhance such procedures by clearly highlighting the bony landmark in real-time [25].

Several imaging techniques are employed to examine TMJ structures. The selection of an appropriate imaging method depends on the patient’s medical history, clinical presentation, the role of radiological evaluation in diagnosis and treatment, and the necessity of minimizing radiation exposure [19, 26–28]. Magnetic resonance imaging (MRI) is considered the gold standard for evaluating the articular disc and soft tissues, whereas cone beam computed tomography (CBCT) provides high-resolution views of osseous components, including the condyle. However, these modalities have limitations: MRI is expensive and not always readily available, and CBCT involves ionizing radiation. In contrast, ultrasonography is a non-invasive, radiation-free, and cost-effective imaging technique that is portable and allows dynamic examination [29, 30].

US is an alternative imaging technique that is non-invasive, cost-effective, and free from ionizing radiation. It uses high-frequency sound waves transmitted by a probe (transducer), which reflects off tissues to generate an image on a screen [30–33]. US has proven valuable in diagnosing TMJ disorders, including disc displacements, joint effusions, and intra-articular abnormalities [33, 34]. Various imaging techniques, such as MRI, CT, and ultrasound, can be used to assess the mandibular condyle [35, 36]. Ultrasound, in particular, is highly effective for visualizing superficial structures in the head and neck region [37–41]. Its affordability and non-ionizing nature make it a convenient and efficient method for evaluating the mandibular condyle [37, 39, 40]. Nevertheless, if the mandibular condyle can be reliably identified and segmented on ultrasound, it could serve as a convenient imaging alternative or adjunct for TMJ assessment, especially in situations where MRI/CBCT are contraindicated or unavailable. This would be beneficial in screening for condylar anomalies or monitoring changes over time without exposing patients to radiation.

Automated TMJ segmentation is the process of automatically identifying and dividing the temporomandibular joint’s borders in medical images using computer algorithms and artificial intelligence techniques. In medical imaging analysis, this procedure is crucial because it facilitates better diagnosis and treatment of illnesses and disorders pertaining to the TMJ. Ito et al. [42] comprised 217 MRI images from individuals with normal or displaced articular discs and segmented TMJ articular discs using deep learning-based semantic segmentation. The study employed three deep learning-based semantic segmentation approaches: SegNet-Basic, U-Net, and 3DiscNet. Of these, SegNet-Basic and 3DiscNet were shown to provide more successful results, with F1 scores of 0.70 and 0.74, respectively. Vinayahalingam et al. [43] used 154 CBCT images in a different investigation with the goal of utilizing the U-Net architecture to segment the glenoid fossa and mandibular condyle. The study’s findings were that the F1 scores for the U-Net architecture’s mandibular condyle and glenoid fossa were 0.976 and 0.966, respectively.

The research on the topic that has been published in the literature has used several imaging modalities; however, no study has yet been done on the segmentation of the mandibular condyle utilizing ultrasound images and deep learning techniques. This study sought to determine how well diagnostic computer software segmented the mandibular condyle—a crucial part of the TMJ—in ultrasound images.

Materıal and Methods

Study Design and Clinical Information

This study was performed in line with the principles of the Declaration of Helsinki. The study protocol of this study was approved by the Kahramanmaraş Sütçü İmam University Medical Research Ethics Committee (TAREK) on 17.02.2025 with session number 2025/06 and decision number 05. USG images of 668 patients admitted to the Kahramanmaraş Sütçü İmam University Faculty of Dentistry between the years 2023 and 2024 were used in this retrospective study. All medical records were anonymised.

The following inclusion and exclusion criteria were applied in this study:

• Age criteria: Individuals younger than 18 or older than 65 were excluded to focus the research on adults with standard TMJ anatomy.

• History of TMJ surgery: Subjects who had undergone prior surgical procedures on the TMJ were excluded, as such interventions could modify joint structure and potentially influence ultrasonographic results.

• Significant TMJ deformities: Participants with congenital or pronounced acquired TMJ abnormalities—such as joint ankylosis—that could hinder precise ultrasound imaging were not included.

• Systemic health conditions: Patients diagnosed with systemic diseases that significantly impact TMJ morphology, including advanced rheumatoid arthritis or severe osteoarthritis with extensive deformities, were excluded from the study.

Image Acquisition and Data Collection

The retrospective assessments were performed using a Fujifilm Arietta 65 (Hitachi, Tokyo, Japan) with an 8-MHz linear active matrix transducer (with a frequency range of 5 to 13 MHz). The ultrasonograms were captured in MM mode with a 3.5-cm image depth and an echo gain of 80–90 dB.

Each of the patients had their heads in the standard position and were sitting upright. Rest was recommended for the patients. The image of the muscle was obtained by scanning the surface of the mandibular ramus at approximately 2.5 cm above the inferior border of the jaw with minimal pressure, and both sides of the masseter muscle were perpendicular to the muscle’s anterior border.

Using the artificial intelligence-powered CranioCatch application (CranioCatch, Eskişehir, Turkey), anonymous retrospective ultrasound images of the mandibular condyle were detected and labeled (Fig. 1).

Fig. 1.

Example demonstrating the use of the CranioCatch (CranioCatch, Eskişehir, Turkey) to identify and label the mandibular condyle in USG images

Ground Truth

Before starting the actual segmentation in the study, three board-certified dentomaxillofacial radiologists (initials: H.Y., 5 years experience; A.G.Ö.T., 6 years; G.K., 12 years) were calibrated to identify and agree on mandibular condyle anatomy. For such objectives, a series of 50 USG images (not included in this study) were examined. First, an initial contour suggestion for the condyle was obtained using the software’s detection feature, which helped standardize the region of interest. Then, three board-certified dentomaxillofacial radiologists (initials: H.Y., 5 years experience; A.G.Ö.T., 6 years; G.K., 12 years) independently adjusted and refined the condyle outlines using a polygonal segmentation tool. The condyle’s bony outline visible on ultrasound was traced, limiting the segmentation to the condylar surface and the immediate 3 mm of subcortical area (to focus on the cortex outline). After independent annotations, a senior maxillofacial radiologist (F.N.P., > 20 years experience) reviewed all segmentations. In cases of discrepancy among the three annotators, the differences were resolved by discussion, and a consensus segmentation mask was finalized for each image. This consensus mask served as the ground truth for model training and evaluation. An example of a labeled ultrasound image is shown in Fig. 1, where the mandibular condyle appears as a bright line (bone surface) delineated by the annotation overlay.

Deep Convolutional Neural Network Architecture

The YOLOv8 (You Only Look Once) architecture was used to carry out the deep learning process (Fig. 2). Convolutional neural networks are used in this one-stage deep learning approach to detect objects. The input module, spine network, neck network, and prediction network are the four primary components of the YOLO architecture. YOLO separates the image into sections and creates boxes with objects in these sections, whereas other algorithms use the entire image to detect objects [44, 45]. The detection process is successfully converted into a regression problem by the one-stage object identification system YOLOv8. Real-time identification and recognition of small objects are made possible by YOLOv8’s superior accuracy and faster detection speed when compared to other methods [46]. The YOLOv8 network may be scaled to create five distinct network models of varying sizes to satisfy the needs of diverse applications. By employing an anchorless strategy rather than the conventional anchor-based technique, YOLOv8 sets itself apart from earlier iterations [46, 47].

Fig. 2.

An example of the entire process for segmenting the mandibular condyle using a deep learning-based method, in this case, YOLOv8

The YOLOv8 architecture was selected for automated segmentation of the mandibular condyle on ultrasound due to its advantages in speed, precision, and object detection capabilities. Unlike traditional segmentation networks such as U-Net, SegNet, or DeepLab, which rely on multi-stage processes, YOLOv8 employs a one-stage architecture that enables both detection and segmentation in a single forward pass, making it well-suited for real-time applications like live ultrasound imaging [48]. Moreover, YOLOv8 has outperformed several contemporary segmentation models in both speed and accuracy across a wide range of object segmentation tasks [45]. While models like U-Net are widely used in medical imaging, including TMJ segmentation on CBCT and MRI [49], they lack explicit object localization capabilities and may produce false positives in irrelevant regions. In contrast, YOLOv8’s object-first segmentation strategy helps minimize such errors by confirming the presence of the condyle before segmentation.

Training Phase

Six hundred sixty-eight anonymised ultrasound images were downsized to 640 × 640 while maintaining the aspect ratio for the segmentation model. The Opencv, Pytorch, Numpy, Pandas, Torch Vision, Torch, Tensorboard, and Seaborn libraries, along with the open-source Python programming language, were used to generate a random dataset. Training details of the deep learning model are given in Table 1. The dataset was divided into three sections, 80% training, 10% validation, and 10% testing, in order to prevent the need for training images for retesting:

• Eighty percent of the photos in the training group correspond to the dataset that was used to train the model.

• Ten percent of the photos in the validation group are unrelated to model training and should not be viewed by the model during this time. In order to assess the model and determine whether to stop training or alter the training variables, this dataset is utilized.

• Ten percent of the images make up the test group, which is used to compare the trained model to both training and validation data.

Table 1.

Trainig details of the deep learning model

Experiment details	Training settings	Optimizer and learning rate	Data augmentation
Task: segmentation Model: YOLOv8x-seg (pretrained)	Epochs: 900 Early stopping (patience): 50 epochs Early stopping Criteria: training stops if no improvement for 50 epochs Batch size: 16 Image size: 640	Optimizer: SGD Initial learning rate (lr0): 0.01 Final learning rate factor (lrf): 0.01	HSV saturation (hsv_s): 0.7 HSV brightness (hsv_v): 0.4 Rotation (degrees): 0.2° Translation (translate): 0.1 Scaling (scale): 0.5 Shear (shear): 0.0 Perspective transform (perspective): 0.0 Vertical flip (flipud): 0.5 Horizontal flip (fliplr): 0.5 Mosaic augmentation (mosaic): 1.0

A total of 668 ultrasonography images of the mandibular condyle were divided into the following:

1. Training group: 536 images

2. Validation group: 66 images

3. Test group: 66 images.

Images were trained with YOLOv8 architecture with 900 epochs (Fig. 3).

Fig. 3.

Examples of training data of the deep learning model

Statistical Analysis

The model’s success was measured using the confusion matrix, an informative table that lists the actual and expected situations (Fig. 4). The confusion matrix in Fig. 4 shows that, out of 66 test cases, the model correctly segmented the condyle in 58 cases (TP), missed it in 7 cases (FN), and produced 2 false detections (FP). Since all images contained a condyle, there were no true negatives. Clinically, this is significant since high precision indicates that when the AI marks a condyle, it is very likely correct, while the high recall ensures that most condyles are successfully detected. Occasional misses can be reviewed and corrected by a radiologist, preserving the tool’s overall clinical utility.

Fig. 4.

Confusion matrix

The AI model’s effectiveness was assessed using the procedures and measurements listed below: true positive (TP), false positive (FP), and false negative (FN) rates were calculated:

• TP: the result where the positive class is accurately predicted by the model

• FP: the result where the positive class is mispredicted by the model

• FN: the result where the negative class is mispredicted by the model

The TP, FP, and FN values were then used to compute the following metrics:

Sensitivity (recall)	TP/(TP + FN)
Precision	TP/(TP + FP)
F1 score	2 TP/(2 TP + FP + FN)
Jaccard index (IoU)	(precision × recall)/ (precision + recall—(precision × recall))
Dice score	2 × (precision × recall)/(precision + recall)

To quantitatively evaluate the model’s performance, we calculated several metrics on the test set comparing the YOLOv8 predictions to the ground truth annotations. The primary metrics were sensitivity (recall), precision, and F1 score, which are standard for object detection and segmentation tasks:

• Sensitivity (recall) = TP/(TP + FN)—the ratio of actual condyles (ground truth) that were correctly detected or segmented by the model. In this context, a TP was counted when the model successfully detected and segmented the condyle in an image (with sufficient overlap to ground truth), a FN when it failed to identify the condyle in an image, and a FP when the model identified a condyle where there was none (or an extra, incorrect region). A high recall means the model misses very few condyles, which is important clinically to avoid overlooking pathology.

• Precision = TP/(TP + FP)—the ratio of condyle detections by the model that were actually correct. High precision indicates few false alarms (the model rarely misidentifies other structures as a condyle). Clinically, this is important to ensure the AI does not confuse other anatomical features or noise with the condyle, which could mislead a diagnosis.

• F1 score = 2·TP/(2·TP + FP + FN)—the harmonic mean of precision and recall. The F1 score provides a single measure of test accuracy that balances both false negatives and false positives. An F1 near 1.0 indicates excellent overall detection performance. In our study, since there is essentially one condyle per image, the F1 score also correlates with the accuracy of identifying that condyle.

In addition to detection metrics, we evaluated segmentation quality using overlap measures between the predicted condyle mask and the ground truth mask:

• Dice similarity coefficient (Dice)—defined as Dice = 2 × (precision × recall)/(precision + recall). This measures the overlap between the predicted condyle region and the expert-annotated condyle region, ranging from 0 (no overlap) to 1 (perfect overlap). It is equivalent to the F1 score computed on a pixel-by-pixel basis for the condyle class. Dice is a particularly relevant metric for medical image segmentation, as it directly reflects how well the shape of the condyle is captured.

• Jaccard index/intersection over union (IoU). This is another overlap metric, closely related to Dice (Dice = 2·IoU/(IoU + 1). IoU is often used in object detection to evaluate localization accuracy. In segmentation, it similarly indicates the fraction of the total area of both the predicted and true condyle that overlaps. An IoU of 0.5, for instance, means 50% overlap. We report mean IoU across the test set.

In YOLO, the evaluation of true positives (TP) and false positives (FP) is performed on an object basis rather than a pixel basis, using the intersection over union (IoU) metric. IoU measures the overlap between the predicted bounding box and the ground truth bounding box. A prediction is considered a true positive if its IoU exceeds a certain threshold (default, 0.50). If the IoU is below the threshold or if a non-existent object is detected, it is classified as a false positive. The ROC curve is a crucial performance metric for classification problems. ROC is a probability curve, and the area under the curve (AUC) represents the degree or measure of separability (Fig. 5).

Fig. 5.

ROC curve for YOLO segmentation

Results

In our study, the artificial intelligence (AI) deep learning model YOLOv8 provided the detection and segmentation of all test images (Fig. 6). Table 2, Figs. 7 and 8 display the AI model estimation performance numbers and graphs for precision, sensitivity, F1 scores, and mandibular condyle label correlation. In Table 2, the F1, precision, and sensitivity scores are 0.93, 0.96, and 0.90 respectively. These numbers show that, on a scale of 0 to 1, the algorithm’s success is near 1.

Fig. 6.

The YOLO v8 algorithm’s segmentation of the mandibular condyle is remarkably near to reality in the prediction of the mandibular condyle (real, manually labeled images; prediction, images detected by the model)

Table 2.

Evaluation for diagnostic performance by AI model set for mandibular condyle segmentation

		Performance value
	Precision	Sensitivity	F1 score	Jaccard index	IoU	Dice score
Mandibular condyle	0.96	0.90	0.93	0.607	0.607	0.736

Fig. 7.

Precision-confidence curve, recall-confidence curve, F1 confidence curve, and precision-recall curve performance graph of the YOLOv8-based AI model

Fig. 8.

To visually compare labels, they are presented as label correlation. The size of the points in the graph represents the confidence score of YOLOv8 for each object. The bigger the point, the higher the confidence score

The precision-confidence curve, recall-confidence curve, F1 confidence curve, and precision-recall curve performance graphs for the YOLOv8-based AI model are shown in Fig. 7. The recall-confidence curve shows the model’s recall score at a given level of confidence. The greatest value of the curve indicates the confidence level at which the model achieves the highest recall score. The recall-confidence curve of the YOLOv8 algorithm also displays the model’s recall score at each confidence level. The curve’s biggest value indicates the confidence level at which the model reaches its highest recall score. The precision-confidence graph indicates that as confidence levels increase, so does precision. However, when the degree of confidence is too high, the precision starts to decrease. This is due to the fact that when the algorithm has a high level of confidence in the accuracy of the objects it detects, it is more likely to generate false positives.

Dıscussıon

TMJ is a critical and complex structure that connects the mandible to the skull, facilitating essential functions such as chewing, swallowing, and speaking. One of its most prominent features is the mandibular condyle, which controls the vertical and sagittal growth of the mandible and can vary in size and shape. Pathological changes to the condyle can occur due to conditions like infections, trauma, neoplasms, condylar hyperplasia, and ankylosis, impacting the overall function of the TMJ [50, 51]. Image segmentation is a vital technique in computer vision, which involves dividing an image into smaller segments or regions with similar visual characteristics. This process enables the extraction of important features and is fundamental to a variety of applications, including medical imaging, object recognition, and autonomous driving. Deep learning (DL) techniques, particularly convolutional neural networks (CNNs), have significantly improved segmentation accuracy by enabling the automatic extraction of complex visual patterns [24, 52, 53]. Despite these advancements, traditional segmentation techniques remain labor-intensive and observer-dependent, particularly in clinical settings. To improve the efficiency and reduce errors in TMJ morphological analysis, there is a need for further automation in segmentation processes [54].

Artificial intelligence, especially deep learning techniques like CNNs, has outperformed conventional segmentation methods by automating the extraction and merging of details in images. Although the potential for CNNs to segment various anatomical features in CBCT images is evident, their application in TMJ segmentation within dentistry remains underexplored. A limited number of studies have utilized deep learning methods for TMJ segmentation, and much of the research on this topic is still in its infancy [54, 55]. One promising deep learning framework for real-time object recognition is the YOLOv8 model, known for its speed and accuracy. YOLOv8 processes an entire image in a single evaluation, making it ideal for real-time analytical applications [46, 50]. Despite the promising capabilities of models like YOLOv8, research specifically targeting the automatic segmentation of the TMJ in CBCT images remains limited [53–55]. For instance, Kim et al. employed a modified U-Net model to automate the assessment of cortical thickness, achieving an intersection over union (IoU) of 0.734 for cortical bone and 0.870 for marrow bone [53]. Similarly, Le et al. obtained a dice score of 0.91 for the mandibular ramus and condyles using a 2D slice-by-slice segmentation method with U-Net [55].

Orhan et al. conducted a study comparing AI-generated STL files to diagnoses made by six radiologists to assess the reliability of these files in identifying osseous alterations of the mandibular condyle. The study found high interclass correlation (ICC) values for conditions such as bifid condyle formation, osteophyte production, and condyle flattening, but low reliability for diagnosing osteosclerosis and erosion. The ICC values for these conditions ranged from 1.000 for flattening and bifid condyle formation to 0.000 for osteosclerosis, indicating that while AI-generated STL files can reliably detect certain mandibular changes, there are still limitations in detecting others, particularly when the severity of erosion influences the accuracy of the diagnosis [52]. These findings highlight the potential of AI in TMJ imaging but also emphasize the need for continued refinement in model performance, especially for more complex conditions.

Lasek et al. [56] developed and validated an AI-based approach for automatically measuring the width of the TMJ space using ultrasound images. They analyzed 142 TMJ ultrasound images, dividing them into three anatomical regions: the mandibular condyle, joint space, and glenoid fossa. Among the tested models, Residual U-Net was chosen as the best option for further refinement, outperforming other models like U-Net + +, SegResNet, V-Net, Attention U-Net, and DeepLabV3 in both accuracy and computational efficiency. A variety of deep learning models were tested, with the 2D Residual U-Net proving to be the most effective in terms of segmentation performance. The model performed well, especially for segmenting the mandibular condyle (Dice, 0.91 ± 0.08), showing strong potential for clinical applications. In our study, which focused solely on the mandibular condyle segmentation and had a larger dataset, the sensitivity and precision scores were found to be 0.90 and 0.96, respectively, yielding results almost identical to this study.

Ultrasound images have distinct challenges (speckle noise, lower contrast, operator-dependent quality) that make automated segmentation difficult, and thus represent a gap in the literature. We selected ultrasound for this study due to its clinical advantages mentioned clearly, and we aimed to fill the gap by exploring whether modern AI techniques can overcome ultrasound’s limitations for TMJ analysis. In particular, we investigated the YOLOv8 architecture for this task. Unlike classical segmentation networks (e.g., U-Net, SegNet, DeepLab) that perform pixel-wise classification over the entire image, YOLOv8 combines object detection with instance segmentation: it simultaneously localizes the object (mandibular condyle) with a bounding box and generates a segmentation mask for that object in a single forward pass. The unified, real-time nature of YOLOv8 is well-suited for ultrasound applications where the condyle must be identified quickly amidst noisy data. By using YOLOv8, we hypothesized that we can achieve robust condyle segmentation with potential for real-time inference during ultrasound examinations—a distinct advantage over slower two-stage or purely segmentation-based models.

The model demonstrated higher precision and recall compared to Dice and IoU scores, suggesting it excels at detecting the mandibular condyle’s location, though its segmentation outlines are not always perfectly accurate. Clinically, this is acceptable, as precise localization is often more critical than exact boundaries. The slightly lower Dice highlights a potential area for refinement, such as post-processing with techniques like conditional random fields or using a secondary refinement network. Alternatively, a hybrid approach—localizing with YOLOv8 and refining with U-Net—could enhance accuracy, though at a computational cost. For real-time applications, YOLOv8 is still a very effective one-time solution.

Our study aims to fill a gap in the literature by developing and testing computer-based diagnostic software for assessing mandibular condyle segmentation, specifically using ultrasound images. Despite promising results, this study has several limitations. Firstly, it focused solely on the mandibular condyle in ultrasound images, without addressing other key TMJ components, such as the articular disc or joint capsule. Segmenting the disc, which appears as a hypo-echoic area during mouth opening, remains a challenge due to its subtle and dynamic nature in ultrasound, and would require specialized labeled datasets. Secondly, the model was trained and tested on data from a single institution using the same ultrasound machine and settings, limiting generalizability. Since ultrasound image quality varies with equipment and operator technique, external validation across different centers is necessary to confirm model robustness, and some fine-tuning may be required. Lastly, while YOLOv8 is capable of real-time analysis, this study was conducted offline on static images. Future work should involve integrating the model into real-time ultrasound workflows to evaluate its impact on clinical performance and diagnostic outcomes.

Conclusıon

The AI-based approach for TMJ ultrasonographic analysis presents notable clinical advantages, particularly in overcoming the challenges related to variability and reproducibility in TMJ imaging. The proposed AI model achieves high precision in segmenting essential anatomical structures, such as the mandibular condyle. By minimizing human variability, the AI-driven method enhances objectivity and reliability, ensuring consistent diagnostic outcomes across diverse clinical environments.

AI plays a crucial role in various aspects of the healthcare industry. The growing workload and potential physician fatigue may compromise diagnostic accuracy and patient outcomes. Integrating artificial intelligence into imaging systems can help alleviate this burden while enhancing efficiency. Our initial findings suggest that deep learning has the potential to address this significant challenge. As AI continues to advance in handling large datasets, future research should prioritize human interests as its core objective.

Author Contribution

Gaye Keser: conceptualization, methodology, writing—original draft preparation. Hakan Yülek: data curation. Ayşe Gül Öner Talmaç: data curation. Filiz Namdar Pekiner: conceptualization, reviewing and editing, supervision. İbrahim Şevki Bayrakdar: software, statistical analysis. Özer Çelik: software, statistical analysis.

Data Availability

Raw data that support the findings of this study are available from the corresponding author, upon reasonable request.

Declarations

Ethics Approval

The study protocol of this study was approved by the Kahramanmaraş Sütçü İmam University Medical Research Ethics Committee (TAREK) on 17.02.2025 with session number 2025/06 and decision number 05.

Conflict of Interest

The authors declare no competing interests.