Panoramic Radiograph–based Deep Learning Models for Diagnosis and Clinical Decision Support of Furcation Lesions in Primary Molars

Panoramic Radiograph–based Deep Learning Models for Diagnosis and Clinical Decision Support of Furcation Lesions in Primary Molars | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Panoramic Radiograph–based Deep Learning Models for Diagnosis and Clinical Decision Support of Furcation Lesions in Primary Molars Nevra Karamüftüoğlu, Ayşe Bulut, Murat Akın, Şeref Sağıroğlu This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7322034/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background: Furcation lesions in primary molars are critical in pediatric dentistry as they often guide decisions between root canal treatment and extraction. This study introduces a novel deep learning-based clinical decision-support system that directly maps radiographic lesion characteristics to corresponding treatment recommendations—a first in the context of pediatric dental imaging. Methods: A total of 387 anonymized panoramic radiographs from children aged 3–13 was labeled into five distinct bone lesion categories. Three object detection models (YOLOv12x, RT-DETR-L, and RT-DETR-X) were trained and evaluated using stratified train-validation-test splits. Diagnostic performance was assessed using precision, recall, [email protected] , and [email protected] –0.95. Additionally, qualitative accuracy was evaluated with expert-annotated samples. Results: Among the models, RT-DETR-X achieved the highest accuracy ( [email protected] = 0.434). All models successfully identified lesion types and supported corresponding clinical decisions. The system reduced diagnostic ambiguity and showed promise in supporting clinicians with varying levels of experience. Conclusions: This study presents the first deep learning–based decision-support framework that links primary molar furcation lesion classification with treatment planning using panoramic radiographs. The proposed models have potential for standardizing diagnostic outcomes, especially in resource-limited settings and mobile clinical environments. AI in dentistry deep learning pediatric endodontics furcation lesion clinical decision-support panoramic radiograph RT-DETR YOLOv12x Figures Figure 1 Figure 2 Figure 3 Background Dental diseases have existed since as early as 2500 BCE and continue to pose a major public health problem today [ 1 ]. Particularly in childhood, caries developing in primary molars may lead to periapical lesions in the furcation areas, and if not diagnosed in time, the infection can spread and result in serious consequences such as damage to the developing permanent teeth [ 2 , 3 ]. In such cases, the correct clinical decision between root canal treatment and tooth extraction is essential for the child’s general health and orofacial development. Artificial intelligence (AI), which mimics human-like cognitive functions through algorithms, has been increasingly applied across various fields of medicine, including dentistry [ 4 , 5 ]. Convolutional neural networks (CNNs), in particular, have demonstrated high accuracy in recognizing and classifying complex patterns in dental radiology, providing more precise results in detecting periapical lesions compared to routine visual observation [ 6 ]. Although advanced imaging methods such as cone-beam computed tomography (CBCT) offer greater sensitivity, their use in daily pediatric dental practice is limited due to disadvantages such as radiation dose, cost, and accessibility. Therefore, AI systems working on two-dimensional (2D) images like panoramic radiographs are considered practical solutions to support diagnosis in pediatric patient groups [ 5 , 7 ]. The complex anatomical and physiological structure of the oral cavity—where mineralized tissues such as dentine and enamel interact with soft tissues—plays a key role in the onset and progression of dental pathologies. Odontoblasts, lining the pulp chamber, act as specialized barrier cells that can detect pathogens, secrete antimicrobial factors, and initiate immune responses to contain the spread of caries-related bacteria toward the periapical tissues [ 8 ]. However, when such defense mechanisms are overwhelmed, bacterial invasion of the pulp and subsequent apical bone tissue may occur, requiring timely clinical intervention. Healing can occur after pathogen removal, which is achieved by disinfection and obturation of the pulp space by root canal treatment. Çelik et al. [ 9 ] demonstrated that a deep learning (DL) model based on panoramic radiographs could detect periapical lesions with high accuracy and emphasized the potential of AI as a clinical decision support system, especially in children and adolescents. Therefore, it is essential to develop AI-based systems in pediatric dentistry to assist clinicians in making accurate treatment decisions, particularly when choosing between root canal therapy and extraction. Unlike previous studies that focused only on lesion detection, we propose a novel DL-based framework that links radiographic diagnosis to treatment decisions. This study aimed to address this specific gap by evaluating a deep learning-based model designed to support clinical decision-making regarding root canal treatment versus extraction in primary molars. Methods In this study, a total of 387 anonymized OPG (Orthopantomogram) images obtained from clinical panoramic X-rays (retrospectively collected from children aged 3 to 13) were first annotated in COCO format according to five bone anomaly classes. Following this, Unicode-based file name normalization and a stratified 70/20/10 train/validation/test split were performed. The implementation steps of the study are illustrated in Figure 1. The images shown in the figure were resized to 512×512 pixels (for YOLOv12x) and 640×640 pixels (for RT-DETR), in accordance with architectural requirements, and underwent augmentation steps. The YOLOv12x, RT-DETR-L, and RT-DETR-X models were fine-tuned using their own COCO pre-trained weights. Training and validation processes were conducted using the Ultralytics API and Transformer-based DETR variants. In the testing phase, quantitative evaluations were carried out using global and class-based Precision, Recall, [email protected] , and [email protected] –0.95 metrics, along with qualitative analyses on selected case samples. This comprehensive model enables the direct transfer of model selection—ensuring both speed-performance balance and reliable detection of subtle bone changes—into clinical decision support systems. 1. Data Collection This retrospective study was approved by the Ethics Committee of Gazi University (Approval No: 2023 - 1264), and all procedures were conducted in accordance with the principles of the Declaration of Helsinki. Written informed consent was obtained from all participants and/or their legal guardians prior to inclusion in the study.The data used in this study consists of 387 panoramic radiographs collected between 2023 and 2024 from the Öveçler Oral and Dental Health Clinic in Ankara. The study included patients aged 3 to 13 who were in good general health and had a routine panoramic radiograph indication due to periapical lesions in their primary molars. Written informed consent was obtained from all participants and/or their legal guardians prior to inclusion in the study. The details regarding the acquisition and processing of the images within the scope of the study are as follows: · Inclusion Criteria: Radiographic detection of periapical lesions in primary molars Availability of high-resolution panoramic OPG images No history of systemic bone diseases · Exclusion Criteria: Severe motion artifacts or missing mandibular borders compromising image quality Presence of post-surgical traces Incomplete DICOM metadata Additionally, radiographic images were selected according to international quality criteria; records with low resolution, artifacts, missing anatomical structures, or image blurring were excluded. · Imaging Protocol: Device: BLUEX Pantos DG XP Panoramic X-ray (50/60 Hz, 115–230 V) Parameters: 60 kV, 10 mA, exposure time: 12 seconds Format: DICOM → anonymization → lossless PNG (0.1 mm/pixel) · Demographic Distribution: Total number of images: 387 Age groups: 3–6 (n=120), 7–10 (n=140), 11–13 (n=127) Gender ratio: 48% female, 52% male · Ethics and Data Security: The research was conducted with the approval of the Ethics Committee of Gazi University. All patient identifiers were removed from DICOM headers, and data was stored in an encrypted format. At this stage, the aim was to maximize data quality and ensure patient privacy before starting model training. 2. Annotation and Labeling Accurate and consistent annotation of the five bone anomaly classes on the panoramic OPGs is crucial for model performance and plays a key role. The procedures followed in this step are as follows: i. The classes used as lamina dura widening, Lamina dura loss, Bone loss one third, Bone loss two thirds, Bone loss three thirds, which the definitions were based on the World Health Organization’s (WHO) alveolar bone loss criteria and periodontal literature. Radiographic characteristics such as margin irregularity, changes in opacity, and gap width were considered as selection criteria. ii. The annotation was done using CVAT, which outputs COCO-compatible JSON files containing class labels. An experienced dentist independently reviewed and annotated each image by drawing bounding box coordinates. A second expert verified the markings and gave final approval. Bounding boxes with Intersection over Union (IoU) ≥ 0.5 were considered equivalent and merged into a single label record. Lower-IoU annotations were resolved through consensus. iii. The dataset was organized in strict COCO JSON format to ensure full compatibility with Ultralytics training pipelines and to facilitate seamless conversion into the text‑based annotation structure required by YOLO. In each JSON file, the images section records the image id, width, and height; the annotations section contains each object’s id, corresponding image_id, category_id, and bounding‑box coordinates as [x, y, w, h]; and the categories section enumerates each class with its id and human‑readable name. iv. To guarantee the highest possible annotation fidelity, a multi‑stage quality control protocol was implemented. Whenever lesion boundaries or class assignments proved ambiguous, the two primary radiologist annotators reconvened with a third expert to refine and standardize the annotation guidelines. Furthermore, a randomly selected 10% subset of the dataset was independently re‑evaluated by a third observer, yielding a consistency rate of at least 95 % for both bounding‑box coordinates and class labels across all lesion categories. As given above, the detailed annotation procedure was planned to maximize the reliability and quality of the training data. To align the annotation with clinically relevant decision-making, each bone anomaly class was also associated with a corresponding treatment implication. Specifically, lamina dura loss and lamina dura widening were considered indicative of the need for root canal treatment. Similarly, bone loss extending to one-third of the interradicular region also signaled root canal indication. In contrast, bone loss involving two-thirds or the entire interradicular area (three-thirds) was labeled as indicative of tooth extraction. These thresholds were determined based on established pediatric endodontic guidelines and expert consensus to ensure that model outputs could be directly integrated into real-world treatment workflows. 3. Normalization and Anonymization These normalization and anonymization steps are aimed at ensuring both cross-platform compatibility and the highest level of patient privacy. Following the annotation phase, two simultaneous operations (two‑phase file normalization and anonymization workflow) were carried out to ensure consistency and privacy in the dataset as: i. First, all original DICOM‑derived PNG and JSON file names underwent Unicode sanitization, in which a Python normalization routine replaced Turkish characters (Ç, Ö, Ş, Ü, Ğ, İ) and whitespace with their ASCII equivalents. Following sanitization, each file was systematically renamed using the pattern “LastName_FirstName_ID.png”, guaranteeing error‑free path resolution under Linux, Windows, and macOS, and simplifying automated script execution. ii. Second, to comply with Data Privacy Act requirements, Pydicom was employed to strip every patient identifier and header metadata field (e.g., PatientName, PatientID) from the original DICOM files. The resulting anonymized images were then assigned randomly generated UUIDs (Universally Unique Identifiers), with the mapping between UUID and original clinical record stored separately in an encrypted registry. Consequently, the downstream COCO JSON annotations—and their YOLO‑formatted text counterparts—included only the image_id and category_id, with no residual personal or clinical information. 4. Data Splitting & Preparation The data splitting process was carried out in a stratified and reproducible manner to reduce the risk of overfitting while evaluating the model’s generalization ability. A reproducible, stratified data‐splitting protocol was adopted to preserve the proportional representation of all five anomaly categories in each subset. First, scikit‑learn’s method was used to stratify based on overall label frequencies [10], and because individual images could carry multiple category_id values, the iterative‑stratification package was employed for true multi‑label stratification. This process yielded a 70 % training set (271 images), 20 % validation set (77 images), and 10 % test set (39 images). To streamline model ingestion, each split produced both a COCO‐formatted JSON file (train.json, val.json, test.json) and a simple text list (train.txt, val.txt, test.txt), with each line of the latter pointing to a single image file. Finally, for initializations, the weights were initialized with the random seed (seed number=42) and fixed throughout to guarantee that the exact same split could be regenerated for future experiments and repeatability of the model proposed. This step also ensured consistent and comparable performance metrics across both training and evaluation stages of the models. The image data used for model training and evaluation were preprocessed to meet the architectural requirements through the following steps: i. All training images underwent resolution‐ and contrast‐preserving preprocessing tailored to each model and robust augmentation to improve generalization. First, input dimensions were standardized: YOLO12 exclusively consumed 512×512 px crops, whereas RT‑DETR-L/X models operated on 640×640 px, with all resizing performed under letterbox padding to preserve the original aspect ratio. Next, raw pixel values were scaled from [0,255] to [0, 1], and channel‑wise normalization ((x-mean)/std) was applied using standard ImageNet mean and standard‑deviation vectors. ii. To enrich the training distribution, a suite of augmentations was applied. Geometric transforms included random horizontal and vertical flips, rotations of up to ±10°, and scaling in the range of 0.8–1.2. Color and contrast augmentations comprised HSV jitter (H ±10 %, S ±30 %, V ±30 %) and CLAHE equalization. In the YOLO12 pipeline, Mosaic (50 % probability, mixing four images) and MixUp (15% probability) were deployed to boost small‐object recall, while the RT‑DETR workflows employed random cropping and padding to diversify local window contexts prior to token down-sampling. iii. All bounding‐box annotations were programmatically updated in lockstep with each augmentation via the Albumentations library. During compound operations (Mosaic/MixUp), annotations from multiple source images were concatenated and then pruned using Non‑Max Suppression (NMS) rules to avoid excessive overlap and maintain class balance. iv. Finally, the train.txt, val.txt, and test.txt lists were regenerated to reflect the augmented dataset splits, and corresponding YAML configuration files were produced to orchestrate the augmented data flow seamlessly during model training. This preprocessing step is critical for improving overall model accuracy and ensuring data diversity, thereby preventing overfitting. 5. AI Model Training All AI model training was carried out using a single NVIDIA A40 GPU with 40 GB of VRAM, within a Python 3.9.23 and PyTorch 2.4.1 environment, running on CUDA 12.4. Training was performed using transfer learning and fine-tuning techniques across three different object detection architectures: YOLOv12-x, RT-DETR-L, and RT-DETR-X. Each experiment was conducted with a batch size of 4 and took approximately 4 hours to complete — around 3.5 hours for YOLO12-x, and about 4.5 hours for RT-DETR-L and RT-DETR-X. YOLO (You Only Look Once) is a single-stage object deep learning detection architecture that balances speed and accuracy at a high level. This model divides the input image into grid cells through a single network pass, obtaining both class and box regression results from each cell. This leads to low latency and high FPS performance. In YOLOv12, the added position perceiver module and FlashAttention-based attention blocks efficiently process both local and global context, resulting in improved performance on small and low-contrast objects [11]. Since lesions in panoramic dental radiographs are often located in fine and low-contrast regions, YOLOv12X was chosen as a first model for its ability to combine fast inference with fine-detail detection [12]. For our implementation, the YOLOv12X Training Parameters were selected as: Initial weights: yolo12x.pt; Input size: 512×512; Epochs: 50; Batch size: 4; Learning rate: initial lr0 = 0.008, cosine decay schedule; Optimizer: AdamW (weight_decay = 0.0005); Augmentation: Mosaic=0.5, MixUp=0.15, HSV jitter= (H 0.1, S 0.3, V 0.3); Early stopping: patience = 50. RT-DETR (Real-Time Detection Transformer) is a high-performance object detection model built on the DETR framework, but redesigned for faster, real-time use. Unlike DETR, which applies global self-attention across the entire image, a method that’s both slow and computationally expensive, RT-DETR improves efficiency by using localized attention. It only performs self-attention within small 16×16 patches and reduces memory use by discarding less important tokens. RT-DETR-L is the second method for our implementation, and a streamlined version of RT-DETR is tailored for complex, high-resolution settings such as panoramic medical images or collaborative robotics. Even though it's lighter in design, it still delivers strong detection accuracy thanks to the same attention mechanisms used in RT-DETR. The model is trained on 640×640 images for 50 epochs with a batch size of 4, using the AdamW optimizer (weight decay = 0.0005) and a cosine-decay learning rate starting at 0.001. It also benefits from data augmentation like random cropping, padding, and light flipping [13,14]. The training parameters of RT-DETR-L were selected as: Initial weights: rtdetr-l.pt; Input size: 640×640; Epochs:50; Batch size:4; Learning rate:lr0 = 0.001, cosine decay; Optimizer: AdamW (weight_decay = 0.0005); Windowed attention window:16×16 tokens; Augmentation: random crop, padding, light flipping. RT-DETR-X, which is the third method, builds on RT-DETR-L by adding a deeper backbone and more attention heads — 8 in total — giving it greater capacity to learn detailed features. It uses larger embedding dimensions and additional transformer layers, which makes it especially effective for detecting fine details in clinical images, such as low-contrast anomalies like lamina dura loss, which are vital for accurate diagnosis. Its training setup is mostly the same as RT-DETR-L, but with an added 3-epoch warm-up phase before applying the cosine decay schedule. Thanks to these improvements, RT-DETR-X achieves higher mean Average Precision (mAP) on tasks requiring fine-grained detection [14]. The training parameters of RT-DETR-X were selected as: Initial weights: rtdetr-x.pt; Input size:640×640; Epochs:50; Batch size:4; Learning rate: lr0 = 0.001, 3-epoch warmup + cosine decay; Optimizer: AdamW (weight_decay = 0.0005); Multi-head attention: 8 heads; Augmentation: same as RT-DETR-L. Results The performance of the three models on the test set is summarized in Table 1 . Table 1 Performance Values of Models AI Model Number of Parameters GFLOPs Time (ms/img) Precision Recall [email protected] mAP@ 0.5–0.95 YOLO12x 59.0 M 198.5 ~ 17 0.442 0.333 0.397 0.163 RTDETR-L 31.99 M 103.5 ~ 12 0.326 0.471 0.326 0.122 RTDETRX 65.48 M 222.5 ~ 24 0.440 0.483 0.434 0.187 The metrics presented in this table are used to comprehensively evaluate object detection models in terms of both computational efficiency and predictive performance. Parameter count, GFLOPs, and inference time (ms/img) reflect the model's computational complexity and suitability for real-time applications. Precision measures the proportion of correctly predicted positive detections among all positive predictions, while recall indicates the model's ability to detect all actual positives. [email protected] evaluates detection accuracy at a fixed IoU threshold of 0.5, commonly used for measuring general detection performance. [email protected] –0.95 (COCO standard) averages precision across multiple IoU thresholds (from 0.5 to 0.95), providing a more stringent and holistic assessment of the model’s localization and classification capabilities. [email protected] : The highest mean Average Precision was observed in RT-DETR-X (0.434), followed by YOLOv12x (0.397), and RT-DETR-L (0.326). Recall & Precision : RT-DETR-X (48.3%) and RT-DETR-L (47.1%) offered similar recall values, while YOLOv12x had a lower recall (33.3%). Precision was nearly identical between YOLOv12x (44.2%) and RT-DETR-X (44.0%), both outperforming RT-DETR-L (32.6%). This indicates that Transformer-based models detect more true positives, while YOLOv12x is more selective, maintaining higher precision. [email protected] –0.95 (COCO standard) : RT-DETR-X (0.187) again achieved the highest score, followed by YOLOv12x (0.163) and RT-DETR-L (0.122). This suggests that RT-DETR-X delivers the most stable overall performance across different IoU thresholds. Based on these results: RT-DETR-X achieved the highest overall mAP by balancing both precision and recall, though it requires more computational resources and has a longer inference time. YOLOv12x provides the best balance between speed and performance, with high precision and reasonably strong mAP values. RT-DETR-L , with the lowest GFLOPs and the fastest inference time (~ 12 ms), is well-suited for real-time applications, albeit with moderate mAP. These quantitative findings serve as guidance for selecting the most appropriate deep learning model for different clinical scenarios. In addition to numerical evaluations, sample-based, case-specific qualitative assessments were also performed. Below, three sample cases are presented as Figs. 1 and 2 . For each case, outputs from YOLOv12x (top), RT-DETR-L (middle), and RT-DETR-X (bottom) were displayed. Green boxes indicate “True Labels” annotated by the expert, while blue boxes represent the model predictions (Pred). In Fig. 2 , three ground‑truth lesions were annotated on the panoramic radiograph: a central “bone loss three thirds” and two “lamina dura widening” regions in the lower molar areas. The YOLO12‑x model correctly detected only the central bone loss lesion (confidence = 0.83), but failed to identify either lamina dura widening (TP = 1, FN = 2, FP = 0). RT‑DETR‑L improved recall by capturing the bone loss (0.77) and one widening (0.44), yet introduced a spurious “lamina dura loss” prediction (TP = 2, FN = 1, FP = 1) and exhibited modest confidence scores (0.26–0.58). In contrast, RT‑DETR‑X achieved perfect coverage—detecting the bone loss (0.77) and both widenings (0.57, 0.32) without any false positives or negatives (TP = 3, FN = 0, FP = 0) - demonstrating superior balance between precision and recall in this challenging case. Figure 3 presents two true lesions: a “bone loss two thirds” in the lower left molar region and a “lamina dura widening” immediately adjacent. YOLO12‑x again excels at the conspicuous lesion, correctly identifying the bone loss (0.57) but missing the subtle widening (TP = 1, FN = 1, FP = 0). RT‑DETR‑L captures the bone loss (0.62) but overpredicts widening, generating multiple false positives (TP = 1, FN = 1, FP = 3), indicating class–label confusion despite high recall. RT‑DETR‑X delivers the most reliable performance here as well, detecting both lesions (bone loss: 0.59; widening: 0.27–0.40) with only one false positive (TP = 2, FN = 0, FP = 1) and consistent confidence levels, underscoring its suitability for exhaustive lesion screening. Table 2 CasebyCase Detection Metrics Model Case TP FN FP Precision Recall YOLO12‑x 1 1 2 0 1.000 0.333 2 1 1 0 1.000 0.500 RT‑DETR‑L 1 2 1 1 0.667 0.667 2 1 1 3 0.250 0.500 RT‑DETR‑X 1 3 0 0 1.000 1.000 2 2 0 1 0.667 1.000 The detection metrics of AI models of sample images are given in Table 2 . Across both cases, RT‑DETR‑X consistently achieves full recall with minimal false positives, making it the most reliable for comprehensive lesion detection. YOLO12‑x offers rapid, high‑precision identification of conspicuous bone losses but underperforms on finer lamina dura changes, suggesting a role as a first‑pass screening tool. RT‑DETR‑L, while delivering fast inference, exhibits class confusion and a high false‑positive rate, indicating its optimal use may lie in preliminary filtering rather than definitive decision support. These findings inform model selection according to clinical priorities—speed, precision, or exhaustive sensitivity—in pediatric panoramic radiograph analysis. Discussion Artificial intellegence has gained significant traction in dentistry, offering transformative potential in diagnosis, treatment planning, and clinical workflow optimization. Among AI methodologies, DL, particularly CNNs, has demonstrated superior performance across several diagnostic domains, including dental caries, periodontal bone loss, vertical root fractures, and periapical lesions. Systematic reviews have reported sensitivity and specificity values for CNN-based caries detection ranging from 0.44–0.86 and 0.85–0.98, respectively, with AUC values typically above 0.84 [ 15 ]. In periapical diagnostics, DL models have achieved strong results even on 2D imaging. For instance, Çelik et al. [ 9 ] reported that a CNN trained on panoramic radiographs achieved an AUC of 0.91 in detecting periapical lesions in pediatric populations, demonstrating that AI tools can maintain high diagnostic accuracy despite intrinsic limitations such as image resolution or anatomical superimposition. Similarly, Boztuna et al. [ 16 ] found precision (~ 0.82), recall (~ 0.77), and F1 scores (~ 0.80) using U²-Net models on panoramic images, underscoring the robustness of carefully trained models in periapical detection. Recent advancements in object detection have introduced transformer-based architectures that outperform traditional CNNs, particularly in tasks requiring nuanced contextual understanding. RT-DETR (Real-Time Detection with Transformers) models leverage attention mechanisms to simultaneously capture global image context and fine-grained local features, allowing for more accurate detection of subtle dental pathologies such as lamina dura widening or early-stage periapical radiolucency. These models decouple object queries from positional encoding, enabling flexible and precise localization even in crowded anatomical areas. Studies by Han et al. [ 17 ] and He et al. [ 18 ] have demonstrated that transformer-based detectors such as DETR and RT-DETR achieve superior performance in medical image tasks by reducing false positives and enhancing boundary delineation. As such, the incorporation of transformer-based vision models holds great promise for future applications in dental radiology, where interpretability and precision are crucial. These findings align with recent evaluations in dental imaging and suggest practical integration strategies. In environments with limited resources, RT-DETR-L offers a computationally efficient screening option. For high-throughput general practices, YOLOv12x’s speed and decent detection performance present a balanced solution. For expert diagnostics or legal documentation, RT-DETR-X offers superior precision and can assist radiologists in detecting early lesions that might otherwise go unnoticed. Furthermore, hybrid workflows—where lightweight models perform initial triage followed by confirmation through high-accuracy models or human radiologists—have been proposed as optimal clinical strategies [ 19 , 20 ]. This study aimed to address a specific gap in pediatric dentistry by evaluating a deep learning-based model designed to support clinical decision-making regarding root canal treatment versus extraction in primary molars. Our preliminary findings suggest that the model can achieve diagnostic accuracy comparable to that reported in previous studies and provide a novel, reliable decision support tool in pediatric clinical settings. These results reinforce the growing role of AI in enhancing diagnostic precision and facilitating treatment planning in complex pediatric cases. Our study expands on this foundation by evaluating three advanced object detection models—YOLOv12x, RT-DETR-L, and RT-DETR-X—for their ability to support clinical decision-making regarding root canal treatment versus extraction in primary molars with suspected furcation involvement. A dataset of 387 well-annotated panoramic radiographs from pediatric patients served as the testing ground, providing a clinically relevant and demographically focused evaluation. The treatment thresholds used in labeling (e.g., lamina dura loss = RCT; complete furcation bone loss = extraction) were clinically grounded to enhance model interpretability and decision-making accuracy. Quantitative performance comparison revealed RT-DETR-X as the top-performing model with [email protected] = 0.434 and mAP@[0.5–0.95] = 0.187, excelling at both pronounced bone loss and subtle lamina dura radiolucency. RT-DETR-L, in contrast, offered the lowest inference time (12 ms/image) and computational load (103.5 GFLOPs), thus ideal for real-time or low-resource settings such as mobile dental units. YOLOv12x emerged as a well-balanced model, with high precision (0.442) and moderate mAP (0.397), making it well-suited for fast chairside applications. Nonetheless, successful implementation of AI in clinical practice depends not only on performance metrics but also on addressing practical and ethical limitations. For instance, access to large and diverse datasets remains essential for reliable model training, yet privacy regulations such as GDPR and HIPAA constrain data sharing, necessitating innovative solutions like data augmentation or federated learning. Moreover, many existing models—including ours—suffer from limited generalizability due to single-center datasets. Variations in patient demographics, image quality, and radiographic equipment across institutions can impact model accuracy. Integrating developmental variables like root morphology and eruption timing may further enhance pediatric-specific diagnostic accuracy. Beyond the scope of pediatric dentistry, the proposed AI-based diagnostic models demonstrate potential utility across broader domains of clinical dental practice. While the present study focused on primary molar furcation lesions, the core methodology—mapping radiographic lesion types to clinical treatment decisions—can be adapted to adult dentition and generalized endodontic assessments. In particular, the reliance on panoramic radiographs rather than advanced imaging such as CBCT makes the system highly suitable for widespread application in general dentistry, especially in settings where access to cone-beam imaging is restricted due to cost, radiation dose, or infrastructural limitations. Furthermore, the integration of AI tools into clinical workflows can help standardize diagnostic outcomes across different dental specialties, including endodontics, pediatric dentistry, and general practice. This is particularly advantageous in low-resource settings or underserved regions, where consistent clinical expertise may not be available. The potential for AI to reduce diagnostic variability among practitioners and streamline workflows is especially relevant in these contexts. Moreover, integrating medical history and patient-specific factors into AI analysis could enable more tailored treatment planning. These broader implications position the current work not only as a contribution to pediatric endodontics but also as a foundational framework for scalable, cross-disciplinary AI integration in dental diagnostics. Unlike previous research focusing solely on lesion detection, this study presents a novel diagnosis-to-treatment mapping framework based on furcation lesion severity, which is rarely integrated into previous AI dental studies. To our knowledge, this is the first application of a deep learning–based model that integrates radiographic lesion classification with direct clinical decision thresholds, specifically linking furcation severity to root canal or extraction protocols. This approach represents a significant step forward in the development of explainable and action-oriented clinical decision support systems in dental radiology. While previous models such as those by Li et al. [ 3 ] and Tian et al. [ 11 ] primarily focused on detecting dental pathologies or localizing lesions using object detection frameworks, they lacked direct integration with treatment recommendation logic. These approaches, although technically advanced, did not extend to actionable clinical decisions, leaving a gap between detection and practice. In contrast, our study bridges this gap by implementing a lesion-to-decision mapping strategy, directly linking radiographic severity patterns to evidence-based treatment choices—specifically, root canal treatment or extraction. This design enhances clinical applicability and moves beyond mere diagnostic support, establishing a foundation for fully integrated AI-driven decision-support systems in pediatric dentistry. This study contributes to the expanding evidence base supporting the integration of AI in pediatric dentistry. The evaluation of multiple advanced DL models demonstrated that AI can effectively assist in clinical decision-making processes, particularly in complex diagnostic scenarios such as determining the need for root canal treatment versus extraction. The findings emphasize the potential for AI to enhance diagnostic consistency, streamline treatment planning, and support clinicians in diverse practice environments. Moving forward, the development of ethically grounded, statistically validated, and clinically interpretable. Finally, AI has the potential to significantly impact the diagnosis and treatment planning of primary molar furcation lesions based on panoramic radiographs. However, its effectiveness depends on overcoming challenges related to data quality, interpretability, regulatory compliance, and integration with clinical workflows. As AI continues to evolve, its role in healthcare will likely expand, but it must be carefully implemented with consideration for the human elements of care and the ethical implications of AI decision-making in medical practice. From a clinical standpoint, the integration of AI-based diagnostic models such as those presented in this study offers several key advantages. By providing automated, lesion-specific treatment suggestions, these systems can significantly reduce the cognitive and time burden placed on clinicians during radiographic evaluation, particularly in complex or ambiguous cases. This has the potential to expedite the decision-making process, minimize diagnostic delays, and standardize care across varying levels of clinical expertise. However, it is essential to recognize that the reliability of such systems hinges on the quality and consistency of the input data—variations in panoramic imaging protocols, resolution, and patient positioning can influence model accuracy. In addition, regulatory and ethical challenges remain central barriers to widespread implementation. Diagnostic AI tools must comply with strict medical standards, and unresolved issues such as algorithmic transparency, potential biases, and legal responsibility must be addressed. As these systems become more autonomous, the importance of aligning technological advancement with ethical safeguards grows ever more critical. In resource-constrained settings—where access to specialists or advanced imaging modalities like CBCT is limited—these models could serve as valuable diagnostic aids, particularly in mobile dental units, school-based programs, or rural clinics. Moreover, their implementation may yield cost-effective benefits by reducing procedural errors, unnecessary referrals, and treatment-related complications. The system’s ability to offer reliable decision support regardless of clinician experience also positions it as a useful tool for training young practitioners, reinforcing clinical decision logic, and promoting evidence-based practices in everyday dental care. Conclusions This study presented for the first time to train transformer-based object detection models—YOLOv12x, RT-DETR-L, and RT-DETR-X— for assisting clinical decision-making between root canal treatment and extraction in primary molars with suspected interradicular (furcation) lesions on panoramic radiographs. Across models, advanced AI achieved high diagnostic accuracy in distinguishing endodontically manageable cases from those warranting extraction, supporting greater decision consistency and shorter evaluation times, particularly for less experienced clinicians. Model-specific analysis indicated that RT-DETR-X provided the strongest overall diagnostic performance, effectively capturing both overt furcation bone loss and subtler radiographic cues (e.g., lamina dura widening). RT-DETR-L delivered rapid inference with low computational demand, suggesting suitability for resource-constrained settings such as mobile clinics and rural practices. YOLOv12x offered a pragmatic balance between speed and accuracy, making it well-suited for chairside screening. These results support an AI-integrated workflow in pediatric dentistry that pairs fast preliminary triage with high-precision confirmatory analysis, provided that deployment adheres to ethical and regulatory principles, including algorithmic transparency, patient-data security, and bias mitigation. Future work should include multicenter, longitudinal evaluations spanning diverse ages and growth-related anatomical variables, prospective clinical trials, real-time chairside integration studies, and external validation across heterogeneous populations and care environments to establish generalizability and clinical utility. Collectively, the findings provide foundational evidence for the responsible, efficient, and clinically impactful adoption of deep learning–based decision support in pediatric endodontic diagnostics. Declarations Abbreviations Not applicable Ethical approval and consent to participate This retrospective study was approved by the Ethics Committee of Gazi University (Approval No: 2023 - 1264), and all procedures were conducted in accordance with the Declaration of Helsinki. Consent for publication Not applicable. Availability of data and materials The dataset used and/or analyzed during the current study are available from the corresponding author on reasonable request. Competing Interests The authors declare that they have no competing interests. Funding This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. Author Contributions NK: Conceptualization, Data curation, Formal analysis, Investigation, Writing – original draft. AB: Data curation, Formal analysis, Investigation, Writing – original draft. MA: Methodology, Software, Formal analysis, Writing – review & editing. ŞS: Methodology, Supervision, Writing – review & editing. Acknowledgements The authors would like to thank the staff of the Öveçler Oral and Dental Health Clinic (Ankara, Türkiye) for their support in data access and imaging coordination. References Fejerskov O, Uribe S, Mariño RJ. Dentistry in a historical perspective and a likely future of the profession. In: Career Paths in Oral Health. Cham, Switzerland: Springer; 2018. p. 3–19. https://doi.org/10.1007/978-3-319-89731-8_1 Broadbent JM, Thomson WM, Williams SM. Does caries in primary teeth predict enamel defects in permanent teeth? A longitudinal study. J Dent Res. 2005;84(3):260–4. https://doi.org/10.1177/154405910508400311 Li L, Yang X, Ju W, Li J, Yang X. Impact of primary molars with periapical disease on permanent successors: A retrospective radiographic study. Heliyon. 2023;9:e15613. https://doi.org/10.1016/j.heliyon.2023.e15613 Shan T, Tay FR, Gu L. Application of artificial intelligence in dentistry. J Dent Res. 2021;100(3):232–44. https://doi.org/10.1177/0022034520979835 Chen YW, Stanley K, Att W. Artificial intelligence in dentistry: Current applications and future perspectives. Quintessence Int. 2020;51(3):248–57. Li CW, Lin SY, Chou HS, Chen TY, Chen YA, Liu SY, et al. Detection of dental apical lesions using CNNs on periapical radiograph. Sensors (Basel). 2021;21(21):7049. https://doi.org/10.3390/s21217049 Song J, Kim Y, Hwang JJ. Deep learning-based detection of apical lesions in panoramic radiographs: A comparative study. J Dent Sci. 2022;17(3):1324–31. https://doi.org/10.1016/j.jds.2021.09.006 Galler KM, Weber M, Korkmaz Y, Widbiller M, Feuerer M. Inflammatory response mechanisms of the dentine–pulp complex and the periapical tissues. Int J Mol Sci. 2021;22(3):1480. https://doi.org/10.3390/ijms22031480 Çelik B, Savaştaer EF, Kaya Hİ, Çelik ME. The role of deep learning for periapical lesion detection on panoramic radiographs. Dentomaxillofac Radiol. 2023;52(7):20230118. https://doi.org/10.1259/dmfr.20230118 Scikit-learn. StratifiedShuffleSplit. Available from: https://scikit-learn.org/stable/modules/generated/sklearn.model_selection.StratifiedShuffleSplit.html [Accessed 27 July 2025]. Tian Y, Ye Q, Doermann D. Yolov12: Attention-centric real-time object detectors. arXiv [Preprint]. 2025. arXiv:2502.12524. https://doi.org/10.48550/arXiv.2502.12524 Ultralytics. YOLOv12 Documentation. Available from: https://docs.ultralytics.com/tr/models/yolo12/ [Accessed 27 July 2025]. Lv H, Xiang Q, Xiao J. Object detection based on improved RT-DETR for human-robot collaboration manufacturing system. In: Proceedings of the Fourth International Conference on Mechanical, Electronics, and Electrical and Automation Control (METMS 2024). Bellingham, WA: SPIE; 2024. Vol. 13163, p. 2144–8. Wei J, Shen W, Hu T, Liu Q, Yu J, Huang J. Dynamic-DETR: Dynamic perception for RT-DETR. In: Proceedings of the 2024 6th International Conference on Robotics and Computer Vision (ICRCV). New York, NY: IEEE; 2024. p. 28–32. Khanagar SB, Al-Ehaideb A, Maganur PC, Vishwanathaiah S, Patil S, Baeshen HA, et al. Developments, application, and performance of artificial intelligence in dentistry–A systematic review. J Dent Sci. 2021;16(2):508–22. https://doi.org/10.1016/j.jds.2020.06.018 Boztuna M, Firincioglulari M, Akkaya N, Orhan K. Segmentation of periapical lesions with automatic deep learning on panoramic radiographs: An artificial intelligence study. BMC Oral Health. 2024;24:1332. https://doi.org/10.1186/s12903-024-03925-6 Han T, Hou S, Gao C, Xu S, Pang J, Gu H, et al. EF-RT-DETR: An efficient focused real-time DETR model for pavement distress detection. J Real-Time Image Process. 2025;22:63. He W, Zhang Y, Xu T, An T, Liang Y, Zhang B. Object detection for medical image analysis: Insights from the RT-DETR model. In: Proceedings of the 2025 International Conference on Artificial Intelligence and Computational Intelligence. New York, NY: IEEE; 2025. p. 415–20. Schwendicke FA, Samek W, Krois J. Artificial intelligence in dentistry: Chances and challenges. J Dent Res. 2020;99(7):769–74. https://doi.org/10.1177/0022034520915714 Gaur M, Pal S, Chaudhuri R, Anto OB. Using artificial intelligence. In: Data Intelligence and Cognitive Informatics: Proceedings of ICDICI 2023. Cham, Switzerland: Springer; 2024. Vol. 327. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7322034","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":503798175,"identity":"45cd77d8-c97f-4679-bbeb-1fabf6e25eb1","order_by":0,"name":"Nevra Karamüftüoğlu","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA4klEQVRIiWNgGAWjYPCCA0DMfAzMZGMnWgsbWxoDQwKQZiZeC48ZWAsDIS3mErmPP3zcc0dOfn7Ptwcff2yT52NmYPzwMQe3FssZ6WaSM549MzY4xrvdcEbCbcM2ZgZmyZnbcGsxOHOMjZnnwOHEDWy826R5Em4zArWwMfPi18L8+c+Bw/Xz23iegbTYE9ZyvI1BmuHA4QSGYzxsIC2JxGhhk+w58Mxww7E0oKfSbie3MTM24/fLYTbmDz8O3JGXbz78TOKDzW3b+e3NBz98xKMFG2BsIE39KBgFo2AUjAIMAABSDU/O9EDXlwAAAABJRU5ErkJggg==","orcid":"","institution":"Health Sciences University","correspondingAuthor":true,"prefix":"","firstName":"Nevra","middleName":"","lastName":"Karamüftüoğlu","suffix":""},{"id":503798176,"identity":"b72e7061-e71b-487a-a98b-a8274f1feda3","order_by":1,"name":"Ayşe Bulut","email":"","orcid":"","institution":"Yozgat Bozok University Physiology","correspondingAuthor":false,"prefix":"","firstName":"Ayşe","middleName":"","lastName":"Bulut","suffix":""},{"id":503798182,"identity":"f7a6e263-6ca4-4c4b-bc37-4c1774070c88","order_by":2,"name":"Murat Akın","email":"","orcid":"","institution":"TUSAŞ - Kazan Vocational School, Gazi University","correspondingAuthor":false,"prefix":"","firstName":"Murat","middleName":"","lastName":"Akın","suffix":""},{"id":503798184,"identity":"0da6c265-401a-415e-bc8d-7d16f128f2f7","order_by":3,"name":"Şeref Sağıroğlu","email":"","orcid":"","institution":"Gazi University","correspondingAuthor":false,"prefix":"","firstName":"Şeref","middleName":"","lastName":"Sağıroğlu","suffix":""}],"badges":[],"createdAt":"2025-08-07 22:08:17","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-7322034/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7322034/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":90298873,"identity":"b113f5af-dce5-46b5-b978-14dfa8569994","added_by":"auto","created_at":"2025-09-01 08:48:35","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":116342,"visible":true,"origin":"","legend":"\u003cp\u003eThe flow diagram of the DL-based dental anomaly detection and decision-making model\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-7322034/v1/0768237137c10245dd338d57.png"},{"id":90300986,"identity":"98af0c72-2013-42cc-a818-aaef2a5f87e8","added_by":"auto","created_at":"2025-09-01 08:56:35","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":747228,"visible":true,"origin":"","legend":"\u003cp\u003eDetection results for each of the three models for a sample patient\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-7322034/v1/a15f0fe041afd1f49fe4abd7.png"},{"id":90300987,"identity":"d8391de5-2659-414a-8e8c-758720735054","added_by":"auto","created_at":"2025-09-01 08:56:35","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":670028,"visible":true,"origin":"","legend":"\u003cp\u003eDetection results for each of the three models for another patient\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-7322034/v1/8c7fb42cd7bd6c36e9f10fba.png"},{"id":91149042,"identity":"7bf0154c-a8ad-4279-9cb6-8209540a6525","added_by":"auto","created_at":"2025-09-12 06:46:44","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2340438,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7322034/v1/885bf0d7-5dda-43b7-aab4-391b6c0a875a.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"\u003cp\u003ePanoramic Radiograph–based Deep Learning Models for Diagnosis and Clinical Decision Support of Furcation Lesions in Primary Molars\u003c/p\u003e","fulltext":[{"header":"Background","content":"\u003cp\u003eDental diseases have existed since as early as 2500 BCE and continue to pose a major public health problem today [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Particularly in childhood, caries developing in primary molars may lead to periapical lesions in the furcation areas, and if not diagnosed in time, the infection can spread and result in serious consequences such as damage to the developing permanent teeth [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. In such cases, the correct clinical decision between root canal treatment and tooth extraction is essential for the child\u0026rsquo;s general health and orofacial development.\u003c/p\u003e\u003cp\u003eArtificial intelligence (AI), which mimics human-like cognitive functions through algorithms, has been increasingly applied across various fields of medicine, including dentistry [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. Convolutional neural networks (CNNs), in particular, have demonstrated high accuracy in recognizing and classifying complex patterns in dental radiology, providing more precise results in detecting periapical lesions compared to routine visual observation [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eAlthough advanced imaging methods such as cone-beam computed tomography (CBCT) offer greater sensitivity, their use in daily pediatric dental practice is limited due to disadvantages such as radiation dose, cost, and accessibility. Therefore, AI systems working on two-dimensional (2D) images like panoramic radiographs are considered practical solutions to support diagnosis in pediatric patient groups [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e].\u003c/p\u003e\u003cp\u003e The complex anatomical and physiological structure of the oral cavity\u0026mdash;where mineralized tissues such as dentine and enamel interact with soft tissues\u0026mdash;plays a key role in the onset and progression of dental pathologies. Odontoblasts, lining the pulp chamber, act as specialized barrier cells that can detect pathogens, secrete antimicrobial factors, and initiate immune responses to contain the spread of caries-related bacteria toward the periapical tissues [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. However, when such defense mechanisms are overwhelmed, bacterial invasion of the pulp and subsequent apical bone tissue may occur, requiring timely clinical intervention. Healing can occur after pathogen removal, which is achieved by disinfection and obturation of the pulp space by root canal treatment.\u003c/p\u003e\u003cp\u003e\u0026Ccedil;elik et al. [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e] demonstrated that a deep learning (DL) model based on panoramic radiographs could detect periapical lesions with high accuracy and emphasized the potential of AI as a clinical decision support system, especially in children and adolescents. Therefore, it is essential to develop AI-based systems in pediatric dentistry to assist clinicians in making accurate treatment decisions, particularly when choosing between root canal therapy and extraction. Unlike previous studies that focused only on lesion detection, we propose a novel DL-based framework that links radiographic diagnosis to treatment decisions. This study aimed to address this specific gap by evaluating a deep learning-based model designed to support clinical decision-making regarding root canal treatment versus extraction in primary molars.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003eIn this study, a total of 387 anonymized OPG (Orthopantomogram) images obtained from clinical panoramic X-rays (retrospectively collected from children aged 3 to 13) were first annotated in COCO format according to five bone anomaly classes. Following this, Unicode-based file name normalization and a stratified 70/20/10 train/validation/test split were performed. The implementation steps of the study are illustrated in Figure 1. The images shown in the figure were resized to 512\u0026times;512 pixels (for YOLOv12x) and 640\u0026times;640 pixels (for RT-DETR), in accordance with architectural requirements, and underwent augmentation steps. The YOLOv12x, RT-DETR-L, and RT-DETR-X models were fine-tuned using their own COCO pre-trained weights. Training and validation processes were conducted using the Ultralytics API and Transformer-based DETR variants. In the testing phase, quantitative evaluations were carried out using global and class-based Precision, Recall, [email protected], and [email protected]\u0026ndash;0.95 metrics, along with qualitative analyses on selected case samples. This comprehensive model enables the direct transfer of model selection\u0026mdash;ensuring both speed-performance balance and reliable detection of subtle bone changes\u0026mdash;into clinical decision support systems.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e1. \u0026nbsp; Data Collection\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis retrospective study was approved by the Ethics Committee of Gazi University (Approval No: 2023 - 1264), and all procedures were conducted in accordance with the principles of the Declaration of Helsinki. Written informed consent was obtained from all participants and/or their legal guardians prior to inclusion in the study.The data used in this study consists of 387 panoramic radiographs collected between 2023 and 2024 from the \u0026Ouml;ve\u0026ccedil;ler Oral and Dental Health Clinic in Ankara. The study included patients aged 3 to 13 who were in good general health and had a routine panoramic radiograph indication due to periapical lesions in their primary molars.\u0026nbsp;Written informed consent was obtained from all participants and/or their legal guardians prior to inclusion in the study.\u0026nbsp;The details regarding the acquisition and processing of the images within the scope of the study are as follows:\u003c/p\u003e\n\u003cp\u003e\u0026middot; \u003cstrong\u003eInclusion Criteria:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eRadiographic detection of periapical lesions in primary molars\u003c/p\u003e\n\u003cp\u003eAvailability of high-resolution panoramic OPG images\u003c/p\u003e\n\u003cp\u003eNo history of systemic bone diseases\u003c/p\u003e\n\u003cp\u003e\u0026middot; \u003cstrong\u003eExclusion Criteria:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eSevere motion artifacts or missing mandibular borders compromising image quality\u003c/p\u003e\n\u003cp\u003ePresence of post-surgical traces\u003c/p\u003e\n\u003cp\u003eIncomplete DICOM metadata\u003c/p\u003e\n\u003cp\u003eAdditionally, radiographic images were selected according to international quality criteria; records with low resolution, artifacts, missing anatomical structures, or image blurring were excluded.\u003c/p\u003e\n\u003cp\u003e\u0026middot; \u003cstrong\u003eImaging Protocol:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eDevice: BLUEX Pantos DG XP Panoramic X-ray (50/60 Hz, 115\u0026ndash;230 V)\u003c/p\u003e\n\u003cp\u003eParameters: 60 kV, 10 mA, exposure time: 12 seconds\u003c/p\u003e\n\u003cp\u003eFormat: DICOM \u0026rarr; anonymization \u0026rarr; lossless PNG (0.1 mm/pixel)\u003c/p\u003e\n\u003cp\u003e\u0026middot; \u003cstrong\u003eDemographic Distribution:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTotal number of images: 387\u003c/p\u003e\n\u003cp\u003eAge groups: 3\u0026ndash;6 (n=120), 7\u0026ndash;10 (n=140), 11\u0026ndash;13 (n=127)\u003c/p\u003e\n\u003cp\u003eGender ratio: 48% female, 52% male\u003c/p\u003e\n\u003cp\u003e\u0026middot; \u003cstrong\u003eEthics and Data Security:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe research was conducted with the approval of the Ethics Committee of Gazi University.\u003c/p\u003e\n\u003cp\u003eAll patient identifiers were removed from DICOM headers, and data was stored in an encrypted format.\u003c/p\u003e\n\u003cp\u003eAt this stage, the aim was to maximize data quality and ensure patient privacy before starting model training.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2. \u0026nbsp; Annotation and Labeling\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAccurate and consistent annotation of the five bone anomaly classes on the panoramic OPGs is crucial for model performance and plays a key role. The procedures followed in this step are as follows:\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ei. \u0026nbsp;\u003c/strong\u003eThe\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003eclasses used as lamina dura widening, Lamina dura loss, Bone loss one third, Bone loss two thirds, Bone loss three thirds, which the definitions were based on the World Health Organization\u0026rsquo;s (WHO) alveolar bone loss criteria and periodontal literature. Radiographic characteristics such as margin irregularity, changes in opacity, and gap width were considered as selection criteria.\u003c/p\u003e\n\u003cp\u003eii. The annotation was done using CVAT, which outputs COCO-compatible JSON files containing class labels. An experienced dentist independently reviewed and annotated each image by drawing bounding box coordinates. A second expert verified the markings and gave final approval. Bounding boxes with Intersection over Union (IoU) \u0026ge; 0.5 were considered equivalent and merged into a single label record. Lower-IoU annotations were resolved through consensus.\u003c/p\u003e\n\u003cp\u003eiii. The dataset was organized in strict COCO JSON format to ensure full compatibility with Ultralytics training pipelines and to facilitate seamless conversion into the text‑based annotation structure required by YOLO. In each JSON file, the images section records the image id, width, and height; the annotations section contains each object\u0026rsquo;s id, corresponding image_id, category_id, and bounding‑box coordinates as [x, y, w, h]; and the categories section enumerates each class with its id and human‑readable name.\u003c/p\u003e\n\u003cp\u003eiv. To guarantee the highest possible annotation fidelity, a multi‑stage quality control protocol was implemented. Whenever lesion boundaries or class assignments proved ambiguous, the two primary radiologist annotators reconvened with a third expert to refine and standardize the annotation guidelines. Furthermore, a randomly selected 10% subset of the dataset was independently re‑evaluated by a third observer, yielding a consistency rate of at least 95 % for both bounding‑box coordinates and class labels across all lesion categories.\u003c/p\u003e\n\u003cp\u003eAs given above, the detailed annotation procedure was planned to maximize the reliability and quality of the training data.\u0026nbsp;To align the annotation with clinically relevant decision-making, each bone anomaly class was also associated with a corresponding treatment implication. Specifically, lamina dura loss and lamina dura widening were considered indicative of the need for root canal treatment. Similarly, bone loss extending to one-third of the interradicular region also signaled root canal indication. In contrast, bone loss involving two-thirds or the entire interradicular area (three-thirds) was labeled as indicative of tooth extraction. These thresholds were determined based on established pediatric endodontic guidelines and expert consensus to ensure that model outputs could be directly integrated into real-world treatment workflows.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3. \u0026nbsp; Normalization and Anonymization\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThese normalization and anonymization steps are aimed at ensuring both cross-platform compatibility and the highest level of patient privacy. Following the annotation phase, two simultaneous operations (two‑phase file normalization and anonymization workflow) were carried out to ensure consistency and privacy in the dataset as:\u0026nbsp;\u003c/p\u003e\n\u003cp\u003ei. \u0026nbsp; \u0026nbsp;First, all original DICOM‑derived PNG and JSON file names underwent Unicode sanitization, in which a Python normalization routine replaced Turkish characters (\u0026Ccedil;, \u0026Ouml;, Ş, \u0026Uuml;, Ğ, İ) and whitespace with their ASCII equivalents. Following sanitization, each file was systematically renamed using the pattern \u0026ldquo;LastName_FirstName_ID.png\u0026rdquo;, guaranteeing error‑free path resolution under Linux, Windows, and macOS, and simplifying automated script execution.\u003c/p\u003e\n\u003cp\u003eii. \u0026nbsp; Second, to comply with Data Privacy Act requirements, Pydicom was employed to strip every patient identifier and header metadata field (e.g., PatientName, PatientID) from the original DICOM files. The resulting anonymized images were then assigned randomly generated UUIDs (Universally Unique Identifiers), with the mapping between UUID and original clinical record stored separately in an encrypted registry. Consequently, the downstream COCO JSON annotations\u0026mdash;and their YOLO‑formatted text counterparts\u0026mdash;included only the image_id and category_id, with no residual personal or clinical information.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4. \u0026nbsp; Data Splitting \u0026amp; Preparation\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe data splitting process was carried out in a stratified and reproducible manner to reduce the risk of overfitting while evaluating the model\u0026rsquo;s generalization ability. A reproducible, stratified data‐splitting protocol was adopted to preserve the proportional representation of all five anomaly categories in each subset. First, scikit‑learn\u0026rsquo;s method was used to stratify based on overall label frequencies [10], and because individual images could carry multiple category_id values, the iterative‑stratification package was employed for true multi‑label stratification. This process yielded a 70 % training set (271 images), 20 % validation set (77 images), and 10 % test set (39 images). To streamline model ingestion, each split produced both a COCO‐formatted JSON file (train.json, val.json, test.json) and a simple text list (train.txt, val.txt, test.txt), with each line of the latter pointing to a single image file. Finally, for initializations, the weights were initialized with the random seed (seed number=42) and fixed throughout to guarantee that the exact same split could be regenerated for future experiments and repeatability of the model proposed. This step also ensured consistent and comparable performance metrics across both training and evaluation stages of the models.\u003c/p\u003e\n\u003cp\u003eThe image data used for model training and evaluation were preprocessed to meet the architectural requirements through the following steps:\u003c/p\u003e\n\u003cp\u003ei. All training images underwent resolution‐ and contrast‐preserving preprocessing tailored to each model and robust augmentation to improve generalization. First, input dimensions were standardized: YOLO12 exclusively consumed 512\u0026times;512 px crops, whereas RT‑DETR-L/X models operated on 640\u0026times;640 px, with all resizing performed under letterbox padding to preserve the original aspect ratio. Next, raw pixel values were scaled from [0,255] to [0, 1], and channel‑wise normalization ((x-mean)/std) was applied using standard ImageNet mean and standard‑deviation vectors.\u003c/p\u003e\n\u003cp\u003eii. To enrich the training distribution, a suite of augmentations was applied. Geometric transforms included random horizontal and vertical flips, rotations of up to \u0026plusmn;10\u0026deg;, and scaling in the range of 0.8\u0026ndash;1.2. Color and contrast augmentations comprised HSV jitter (H \u0026plusmn;10 %, S \u0026plusmn;30 %, V \u0026plusmn;30 %) and CLAHE equalization. In the YOLO12 pipeline, Mosaic (50 % probability, mixing four images) and MixUp (15% probability) were deployed to boost small‐object recall, while the RT‑DETR workflows employed random cropping and padding to diversify local window contexts prior to token down-sampling.\u003c/p\u003e\n\u003cp\u003eiii. All bounding‐box annotations were programmatically updated in lockstep with each augmentation via the Albumentations library. During compound operations (Mosaic/MixUp), annotations from multiple source images were concatenated and then pruned using Non‑Max Suppression (NMS) rules to avoid excessive overlap and maintain class balance.\u003c/p\u003e\n\u003cp\u003eiv. Finally, the train.txt, val.txt, and test.txt lists were regenerated to reflect the augmented dataset splits, and corresponding YAML configuration files were produced to orchestrate the augmented data flow seamlessly during model training. This preprocessing step is critical for improving overall model accuracy and ensuring data diversity, thereby preventing overfitting.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e5. \u0026nbsp; AI Model Training\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll AI model training was carried out using a single NVIDIA A40 GPU with 40 GB of VRAM, within a Python 3.9.23 and PyTorch 2.4.1 environment, running on CUDA 12.4. Training was performed using transfer learning and fine-tuning techniques across three different object detection architectures: YOLOv12-x, RT-DETR-L, and RT-DETR-X. Each experiment was conducted with a batch size of 4 and took approximately 4 hours to complete \u0026mdash; around 3.5 hours for YOLO12-x, and about 4.5 hours for RT-DETR-L and RT-DETR-X.\u003c/p\u003e\n\u003cp\u003eYOLO (You Only Look Once) is a single-stage object deep learning detection architecture that balances speed and accuracy at a high level. \u0026nbsp;This model divides the input image into grid cells through a single network pass, obtaining both class and box regression results from each cell. This leads to low latency and high FPS performance. In YOLOv12, the added \u003cem\u003eposition perceiver\u003c/em\u003e module and FlashAttention-based attention blocks efficiently process both local and global context, resulting in improved performance on small and low-contrast objects [11]. Since lesions in panoramic dental radiographs are often located in fine and low-contrast regions, YOLOv12X\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003ewas chosen as a first model for its ability to combine fast inference with fine-detail detection [12].\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;For our implementation, the YOLOv12X Training Parameters were selected as: Initial weights: yolo12x.pt; Input size: 512\u0026times;512; Epochs: 50; Batch size: 4; Learning rate: initial lr0 = 0.008, cosine decay schedule; Optimizer: AdamW (weight_decay = 0.0005); Augmentation: Mosaic=0.5, MixUp=0.15, HSV jitter= (H 0.1, S 0.3, V 0.3); Early stopping: patience = 50.\u003c/p\u003e\n\u003cp\u003eRT-DETR (Real-Time Detection Transformer) is a high-performance object detection model built on the DETR framework, but redesigned for faster, real-time use. Unlike DETR, which applies global self-attention across the entire image, a method that\u0026rsquo;s both slow and computationally expensive, RT-DETR improves efficiency by using localized attention. It only performs self-attention within small 16\u0026times;16 patches and reduces memory use by discarding less important tokens.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eRT-DETR-L is the second method for our implementation, and a streamlined version of RT-DETR is tailored for complex, high-resolution settings such as panoramic medical images or collaborative robotics. Even though it\u0026apos;s lighter in design, it still delivers strong detection accuracy thanks to the same attention mechanisms used in RT-DETR. The model is trained on 640\u0026times;640 images for 50 epochs with a batch size of 4, using the AdamW optimizer (weight decay = 0.0005) and a cosine-decay learning rate starting at 0.001. It also benefits from data augmentation like random cropping, padding, and light flipping [13,14].\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe training parameters of RT-DETR-L were selected as: Initial weights: rtdetr-l.pt; Input size: 640\u0026times;640; Epochs:50; Batch size:4; Learning rate:lr0 = 0.001, cosine decay; Optimizer: AdamW (weight_decay = 0.0005); Windowed attention window:16\u0026times;16 tokens; Augmentation: random crop, padding, light flipping.\u003c/p\u003e\n\u003cp\u003eRT-DETR-X, which is the third method, builds on RT-DETR-L by adding a deeper backbone and more attention heads \u0026mdash; 8 in total \u0026mdash; giving it greater capacity to learn detailed features. It uses larger embedding dimensions and additional transformer layers, which makes it especially effective for detecting fine details in clinical images, such as low-contrast anomalies like lamina dura loss, which are vital for accurate diagnosis. Its training setup is mostly the same as RT-DETR-L, but with an added 3-epoch warm-up phase before applying the cosine decay schedule. Thanks to these improvements, RT-DETR-X achieves higher mean Average Precision (mAP) on tasks requiring fine-grained detection [14].\u003c/p\u003e\n\u003cp\u003eThe training parameters of RT-DETR-X were selected as: Initial weights: rtdetr-x.pt; Input size:640\u0026times;640; Epochs:50; Batch size:4; Learning rate: lr0 = 0.001, 3-epoch warmup + cosine decay; Optimizer: AdamW (weight_decay = 0.0005); Multi-head attention: 8 heads; Augmentation: same as RT-DETR-L.\u0026nbsp;\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eThe performance of the three models on the test set is summarized in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003ePerformance Values of Models\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"8\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAI Model\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eNumber of Parameters\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eGFLOPs\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTime (ms/img)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003ePrecision\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eRecall\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\[email protected]\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u003cp\u003emAP@ 0.5\u0026ndash;0.95\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eYOLO12x\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e59.0 M\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e198.5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e~\u0026thinsp;17\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.442\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.333\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.397\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.163\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRTDETR-L\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e31.99 M\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e103.5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e~\u0026thinsp;12\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.326\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.471\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.326\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.122\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRTDETRX\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e65.48 M\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e222.5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e~\u0026thinsp;24\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.440\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.483\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.434\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.187\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eThe metrics presented in this table are used to comprehensively evaluate object detection models in terms of both computational efficiency and predictive performance. Parameter count, GFLOPs, and inference time (ms/img) reflect the model's computational complexity and suitability for real-time applications. Precision measures the proportion of correctly predicted positive detections among all positive predictions, while recall indicates the model's ability to detect all actual positives. [email protected] evaluates detection accuracy at a fixed IoU threshold of 0.5, commonly used for measuring general detection performance. [email protected]\u0026ndash;0.95 (COCO standard) averages precision across multiple IoU thresholds (from 0.5 to 0.95), providing a more stringent and holistic assessment of the model\u0026rsquo;s localization and classification capabilities.\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\[email protected]\u003c/b\u003e: The highest mean Average Precision was observed in RT-DETR-X (0.434), followed by YOLOv12x (0.397), and RT-DETR-L (0.326).\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eRecall \u0026amp; Precision\u003c/b\u003e: RT-DETR-X (48.3%) and RT-DETR-L (47.1%) offered similar recall values, while YOLOv12x had a lower recall (33.3%). Precision was nearly identical between YOLOv12x (44.2%) and RT-DETR-X (44.0%), both outperforming RT-DETR-L (32.6%). This indicates that Transformer-based models detect more true positives, while YOLOv12x is more selective, maintaining higher precision.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\[email protected]\u0026ndash;0.95 (COCO standard)\u003c/b\u003e: RT-DETR-X (0.187) again achieved the highest score, followed by YOLOv12x (0.163) and RT-DETR-L (0.122). This suggests that RT-DETR-X delivers the most stable overall performance across different IoU thresholds.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\n\u003ch3\u003eBased on these results:\u003c/h3\u003e\n\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eRT-DETR-X\u003c/b\u003e achieved the highest overall mAP by balancing both precision and recall, though it requires more computational resources and has a longer inference time.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eYOLOv12x\u003c/b\u003e provides the best balance between speed and performance, with high precision and reasonably strong mAP values.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eRT-DETR-L\u003c/b\u003e, with the lowest GFLOPs and the fastest inference time (~\u0026thinsp;12 ms), is well-suited for real-time applications, albeit with moderate mAP.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eThese quantitative findings serve as guidance for selecting the most appropriate deep learning model for different clinical scenarios. In addition to numerical evaluations, sample-based, case-specific qualitative assessments were also performed.\u003c/p\u003e\u003cp\u003eBelow, three sample cases are presented as Figs.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e and \u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. For each case, outputs from YOLOv12x (top), RT-DETR-L (middle), and RT-DETR-X (bottom) were displayed. Green boxes indicate \u0026ldquo;True Labels\u0026rdquo; annotated by the expert, while blue boxes represent the model predictions (Pred).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eIn Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, three ground‑truth lesions were annotated on the panoramic radiograph: a central \u0026ldquo;bone loss three thirds\u0026rdquo; and two \u0026ldquo;lamina dura widening\u0026rdquo; regions in the lower molar areas. The YOLO12‑x model correctly detected only the central bone loss lesion (confidence\u0026thinsp;=\u0026thinsp;0.83), but failed to identify either lamina dura widening (TP\u0026thinsp;=\u0026thinsp;1, FN\u0026thinsp;=\u0026thinsp;2, FP\u0026thinsp;=\u0026thinsp;0). RT‑DETR‑L improved recall by capturing the bone loss (0.77) and one widening (0.44), yet introduced a spurious \u0026ldquo;lamina dura loss\u0026rdquo; prediction (TP\u0026thinsp;=\u0026thinsp;2, FN\u0026thinsp;=\u0026thinsp;1, FP\u0026thinsp;=\u0026thinsp;1) and exhibited modest confidence scores (0.26\u0026ndash;0.58). In contrast, RT‑DETR‑X achieved perfect coverage\u0026mdash;detecting the bone loss (0.77) and both widenings (0.57, 0.32) without any false positives or negatives (TP\u0026thinsp;=\u0026thinsp;3, FN\u0026thinsp;=\u0026thinsp;0, FP\u0026thinsp;=\u0026thinsp;0) - demonstrating superior balance between precision and recall in this challenging case.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eFigure \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e presents two true lesions: a \u0026ldquo;bone loss two thirds\u0026rdquo; in the lower left molar region and a \u0026ldquo;lamina dura widening\u0026rdquo; immediately adjacent. YOLO12‑x again excels at the conspicuous lesion, correctly identifying the bone loss (0.57) but missing the subtle widening (TP\u0026thinsp;=\u0026thinsp;1, FN\u0026thinsp;=\u0026thinsp;1, FP\u0026thinsp;=\u0026thinsp;0). RT‑DETR‑L captures the bone loss (0.62) but overpredicts widening, generating multiple false positives (TP\u0026thinsp;=\u0026thinsp;1, FN\u0026thinsp;=\u0026thinsp;1, FP\u0026thinsp;=\u0026thinsp;3), indicating class\u0026ndash;label confusion despite high recall. RT‑DETR‑X delivers the most reliable performance here as well, detecting both lesions (bone loss: 0.59; widening: 0.27\u0026ndash;0.40) with only one false positive (TP\u0026thinsp;=\u0026thinsp;2, FN\u0026thinsp;=\u0026thinsp;0, FP\u0026thinsp;=\u0026thinsp;1) and consistent confidence levels, underscoring its suitability for exhaustive lesion screening.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eCasebyCase Detection Metrics\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"7\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eModel\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eCase\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eTP\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eFN\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eFP\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003ePrecision\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003eRecall\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eYOLO12‑x\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e1.000\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.333\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e1.000\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.500\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRT‑DETR‑L\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.667\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.667\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.250\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.500\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRT‑DETR‑X\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e1.000\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e1.000\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.667\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e1.000\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eThe detection metrics of AI models of sample images are given in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. Across both cases, RT‑DETR‑X consistently achieves full recall with minimal false positives, making it the most reliable for comprehensive lesion detection. YOLO12‑x offers rapid, high‑precision identification of conspicuous bone losses but underperforms on finer lamina dura changes, suggesting a role as a first‑pass screening tool. RT‑DETR‑L, while delivering fast inference, exhibits class confusion and a high false‑positive rate, indicating its optimal use may lie in preliminary filtering rather than definitive decision support. These findings inform model selection according to clinical priorities\u0026mdash;speed, precision, or exhaustive sensitivity\u0026mdash;in pediatric panoramic radiograph analysis.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eArtificial intellegence has gained significant traction in dentistry, offering transformative potential in diagnosis, treatment planning, and clinical workflow optimization. Among AI methodologies, DL, particularly CNNs, has demonstrated superior performance across several diagnostic domains, including dental caries, periodontal bone loss, vertical root fractures, and periapical lesions. Systematic reviews have reported sensitivity and specificity values for CNN-based caries detection ranging from 0.44\u0026ndash;0.86 and 0.85\u0026ndash;0.98, respectively, with AUC values typically above 0.84 [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eIn periapical diagnostics, DL models have achieved strong results even on 2D imaging. For instance, \u0026Ccedil;elik et al. [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e] reported that a CNN trained on panoramic radiographs achieved an AUC of 0.91 in detecting periapical lesions in pediatric populations, demonstrating that AI tools can maintain high diagnostic accuracy despite intrinsic limitations such as image resolution or anatomical superimposition. Similarly, Boztuna et al. [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e] found precision (~\u0026thinsp;0.82), recall (~\u0026thinsp;0.77), and F1 scores (~\u0026thinsp;0.80) using U\u0026sup2;-Net models on panoramic images, underscoring the robustness of carefully trained models in periapical detection.\u003c/p\u003e\u003cp\u003eRecent advancements in object detection have introduced transformer-based architectures that outperform traditional CNNs, particularly in tasks requiring nuanced contextual understanding. RT-DETR (Real-Time Detection with Transformers) models leverage attention mechanisms to simultaneously capture global image context and fine-grained local features, allowing for more accurate detection of subtle dental pathologies such as lamina dura widening or early-stage periapical radiolucency. These models decouple object queries from positional encoding, enabling flexible and precise localization even in crowded anatomical areas. Studies by Han et al. [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e] and He et al. [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e] have demonstrated that transformer-based detectors such as DETR and RT-DETR achieve superior performance in medical image tasks by reducing false positives and enhancing boundary delineation.\u003c/p\u003e\u003cp\u003eAs such, the incorporation of transformer-based vision models holds great promise for future applications in dental radiology, where interpretability and precision are crucial. These findings align with recent evaluations in dental imaging and suggest practical integration strategies. In environments with limited resources, RT-DETR-L offers a computationally efficient screening option. For high-throughput general practices, YOLOv12x\u0026rsquo;s speed and decent detection performance present a balanced solution. For expert diagnostics or legal documentation, RT-DETR-X offers superior precision and can assist radiologists in detecting early lesions that might otherwise go unnoticed. Furthermore, hybrid workflows\u0026mdash;where lightweight models perform initial triage followed by confirmation through high-accuracy models or human radiologists\u0026mdash;have been proposed as optimal clinical strategies [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eThis study aimed to address a specific gap in pediatric dentistry by evaluating a deep learning-based model designed to support clinical decision-making regarding root canal treatment versus extraction in primary molars. Our preliminary findings suggest that the model can achieve diagnostic accuracy comparable to that reported in previous studies and provide a novel, reliable decision support tool in pediatric clinical settings. These results reinforce the growing role of AI in enhancing diagnostic precision and facilitating treatment planning in complex pediatric cases.\u003c/p\u003e\u003cp\u003eOur study expands on this foundation by evaluating three advanced object detection models\u0026mdash;YOLOv12x, RT-DETR-L, and RT-DETR-X\u0026mdash;for their ability to support clinical decision-making regarding root canal treatment versus extraction in primary molars with suspected furcation involvement. A dataset of 387 well-annotated panoramic radiographs from pediatric patients served as the testing ground, providing a clinically relevant and demographically focused evaluation. The treatment thresholds used in labeling (e.g., lamina dura loss\u0026thinsp;=\u0026thinsp;RCT; complete furcation bone loss\u0026thinsp;=\u0026thinsp;extraction) were clinically grounded to enhance model interpretability and decision-making accuracy.\u003c/p\u003e\u003cp\u003eQuantitative performance comparison revealed RT-DETR-X as the top-performing model with [email protected]\u0026thinsp;=\u0026thinsp;0.434 and mAP@[0.5\u0026ndash;0.95]\u0026thinsp;=\u0026thinsp;0.187, excelling at both pronounced bone loss and subtle lamina dura radiolucency. RT-DETR-L, in contrast, offered the lowest inference time (12 ms/image) and computational load (103.5 GFLOPs), thus ideal for real-time or low-resource settings such as mobile dental units. YOLOv12x emerged as a well-balanced model, with high precision (0.442) and moderate mAP (0.397), making it well-suited for fast chairside applications.\u003c/p\u003e\u003cp\u003eNonetheless, successful implementation of AI in clinical practice depends not only on performance metrics but also on addressing practical and ethical limitations. For instance, access to large and diverse datasets remains essential for reliable model training, yet privacy regulations such as GDPR and HIPAA constrain data sharing, necessitating innovative solutions like data augmentation or federated learning.\u003c/p\u003e\u003cp\u003eMoreover, many existing models\u0026mdash;including ours\u0026mdash;suffer from limited generalizability due to single-center datasets. Variations in patient demographics, image quality, and radiographic equipment across institutions can impact model accuracy. Integrating developmental variables like root morphology and eruption timing may further enhance pediatric-specific diagnostic accuracy.\u003c/p\u003e\u003cp\u003eBeyond the scope of pediatric dentistry, the proposed AI-based diagnostic models demonstrate potential utility across broader domains of clinical dental practice. While the present study focused on primary molar furcation lesions, the core methodology\u0026mdash;mapping radiographic lesion types to clinical treatment decisions\u0026mdash;can be adapted to adult dentition and generalized endodontic assessments. In particular, the reliance on panoramic radiographs rather than advanced imaging such as CBCT makes the system highly suitable for widespread application in general dentistry, especially in settings where access to cone-beam imaging is restricted due to cost, radiation dose, or infrastructural limitations.\u003c/p\u003e\u003cp\u003eFurthermore, the integration of AI tools into clinical workflows can help standardize diagnostic outcomes across different dental specialties, including endodontics, pediatric dentistry, and general practice. This is particularly advantageous in low-resource settings or underserved regions, where consistent clinical expertise may not be available. The potential for AI to reduce diagnostic variability among practitioners and streamline workflows is especially relevant in these contexts. Moreover, integrating medical history and patient-specific factors into AI analysis could enable more tailored treatment planning. These broader implications position the current work not only as a contribution to pediatric endodontics but also as a foundational framework for scalable, cross-disciplinary AI integration in dental diagnostics. Unlike previous research focusing solely on lesion detection, this study presents a novel diagnosis-to-treatment mapping framework based on furcation lesion severity, which is rarely integrated into previous AI dental studies. To our knowledge, this is the first application of a deep learning\u0026ndash;based model that integrates radiographic lesion classification with direct clinical decision thresholds, specifically linking furcation severity to root canal or extraction protocols. This approach represents a significant step forward in the development of explainable and action-oriented clinical decision support systems in dental radiology.\u003c/p\u003e\u003cp\u003eWhile previous models such as those by Li et al. [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e] and Tian et al. [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e] primarily focused on detecting dental pathologies or localizing lesions using object detection frameworks, they lacked direct integration with treatment recommendation logic. These approaches, although technically advanced, did not extend to actionable clinical decisions, leaving a gap between detection and practice. In contrast, our study bridges this gap by implementing a lesion-to-decision mapping strategy, directly linking radiographic severity patterns to evidence-based treatment choices\u0026mdash;specifically, root canal treatment or extraction. This design enhances clinical applicability and moves beyond mere diagnostic support, establishing a foundation for fully integrated AI-driven decision-support systems in pediatric dentistry.\u003c/p\u003e\u003cp\u003eThis study contributes to the expanding evidence base supporting the integration of AI in pediatric dentistry. The evaluation of multiple advanced DL models demonstrated that AI can effectively assist in clinical decision-making processes, particularly in complex diagnostic scenarios such as determining the need for root canal treatment versus extraction. The findings emphasize the potential for AI to enhance diagnostic consistency, streamline treatment planning, and support clinicians in diverse practice environments. Moving forward, the development of ethically grounded, statistically validated, and clinically interpretable. Finally, AI has the potential to significantly impact the diagnosis and treatment planning of primary molar furcation lesions based on panoramic radiographs. However, its effectiveness depends on overcoming challenges related to data quality, interpretability, regulatory compliance, and integration with clinical workflows. As AI continues to evolve, its role in healthcare will likely expand, but it must be carefully implemented with consideration for the human elements of care and the ethical implications of AI decision-making in medical practice.\u003c/p\u003e\u003cp\u003eFrom a clinical standpoint, the integration of AI-based diagnostic models such as those presented in this study offers several key advantages. By providing automated, lesion-specific treatment suggestions, these systems can significantly reduce the cognitive and time burden placed on clinicians during radiographic evaluation, particularly in complex or ambiguous cases. This has the potential to expedite the decision-making process, minimize diagnostic delays, and standardize care across varying levels of clinical expertise. However, it is essential to recognize that the reliability of such systems hinges on the quality and consistency of the input data\u0026mdash;variations in panoramic imaging protocols, resolution, and patient positioning can influence model accuracy.\u003c/p\u003e\u003cp\u003eIn addition, regulatory and ethical challenges remain central barriers to widespread implementation. Diagnostic AI tools must comply with strict medical standards, and unresolved issues such as algorithmic transparency, potential biases, and legal responsibility must be addressed. As these systems become more autonomous, the importance of aligning technological advancement with ethical safeguards grows ever more critical.\u003c/p\u003e\u003cp\u003eIn resource-constrained settings\u0026mdash;where access to specialists or advanced imaging modalities like CBCT is limited\u0026mdash;these models could serve as valuable diagnostic aids, particularly in mobile dental units, school-based programs, or rural clinics. Moreover, their implementation may yield cost-effective benefits by reducing procedural errors, unnecessary referrals, and treatment-related complications. The system\u0026rsquo;s ability to offer reliable decision support regardless of clinician experience also positions it as a useful tool for training young practitioners, reinforcing clinical decision logic, and promoting evidence-based practices in everyday dental care.\u003c/p\u003e"},{"header":"Conclusions","content":"\u003cp\u003eThis study presented for the first time to train transformer-based object detection models\u0026mdash;YOLOv12x, RT-DETR-L, and RT-DETR-X\u0026mdash; for assisting clinical decision-making between root canal treatment and extraction in primary molars with suspected interradicular (furcation) lesions on panoramic radiographs. Across models, advanced AI achieved high diagnostic accuracy in distinguishing endodontically manageable cases from those warranting extraction, supporting greater decision consistency and shorter evaluation times, particularly for less experienced clinicians.\u003c/p\u003e\u003cp\u003eModel-specific analysis indicated that RT-DETR-X provided the strongest overall diagnostic performance, effectively capturing both overt furcation bone loss and subtler radiographic cues (e.g., lamina dura widening). RT-DETR-L delivered rapid inference with low computational demand, suggesting suitability for resource-constrained settings such as mobile clinics and rural practices. YOLOv12x offered a pragmatic balance between speed and accuracy, making it well-suited for chairside screening.\u003c/p\u003e\u003cp\u003e These results support an AI-integrated workflow in pediatric dentistry that pairs fast preliminary triage with high-precision confirmatory analysis, provided that deployment adheres to ethical and regulatory principles, including algorithmic transparency, patient-data security, and bias mitigation. Future work should include multicenter, longitudinal evaluations spanning diverse ages and growth-related anatomical variables, prospective clinical trials, real-time chairside integration studies, and external validation across heterogeneous populations and care environments to establish generalizability and clinical utility. Collectively, the findings provide foundational evidence for the responsible, efficient, and clinically impactful adoption of deep learning\u0026ndash;based decision support in pediatric endodontic diagnostics.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003eAbbreviations\u003c/p\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e\n\u003cp\u003eEthical\u0026nbsp;approval and consent to participate\u003c/p\u003e\n\u003cp\u003eThis retrospective study was approved by the Ethics Committee of Gazi University (Approval No: 2023 - 1264), and all procedures were conducted in accordance with the Declaration of Helsinki.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe dataset used and/or analyzed during the current study are available from the corresponding author on reasonable request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting Interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e\n\u003cp\u003eFunding\u003c/p\u003e\n\u003cp\u003eThis research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.\u003c/p\u003e\n\u003cp\u003eAuthor Contributions\u003c/p\u003e\n\u003cp\u003eNK: Conceptualization, Data curation, Formal analysis, Investigation, Writing \u0026ndash; original draft.\u003c/p\u003e\n\u003cp\u003eAB: Data curation, Formal analysis, Investigation, Writing \u0026ndash; original draft.\u003c/p\u003e\n\u003cp\u003eMA: Methodology, Software, Formal analysis, Writing \u0026ndash; review \u0026amp; editing.\u003c/p\u003e\n\u003cp\u003eŞS: Methodology, Supervision, Writing \u0026ndash; review \u0026amp; editing.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors would like to thank the staff of the \u0026Ouml;ve\u0026ccedil;ler Oral and Dental Health Clinic (Ankara, T\u0026uuml;rkiye) for their support in data access and imaging coordination. \u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eFejerskov O, Uribe S, Mari\u0026ntilde;o RJ. Dentistry in a historical perspective and a likely future of the profession. In: Career Paths in Oral Health. Cham, Switzerland: Springer; 2018. p. 3\u0026ndash;19. https://doi.org/10.1007/978-3-319-89731-8_1\u003c/li\u003e\n \u003cli\u003eBroadbent JM, Thomson WM, Williams SM. Does caries in primary teeth predict enamel defects in permanent teeth? A longitudinal study. J Dent Res. 2005;84(3):260\u0026ndash;4. https://doi.org/10.1177/154405910508400311\u003c/li\u003e\n \u003cli\u003eLi L, Yang X, Ju W, Li J, Yang X. Impact of primary molars with periapical disease on permanent successors: A retrospective radiographic study. Heliyon. 2023;9:e15613. https://doi.org/10.1016/j.heliyon.2023.e15613\u003c/li\u003e\n \u003cli\u003eShan T, Tay FR, Gu L. Application of artificial intelligence in dentistry. J Dent Res. 2021;100(3):232\u0026ndash;44. https://doi.org/10.1177/0022034520979835\u003c/li\u003e\n \u003cli\u003eChen YW, Stanley K, Att W. Artificial intelligence in dentistry: Current applications and future perspectives. Quintessence Int. 2020;51(3):248\u0026ndash;57.\u003c/li\u003e\n \u003cli\u003eLi CW, Lin SY, Chou HS, Chen TY, Chen YA, Liu SY, et al. Detection of dental apical lesions using CNNs on periapical radiograph. Sensors (Basel). 2021;21(21):7049. https://doi.org/10.3390/s21217049\u003c/li\u003e\n \u003cli\u003eSong J, Kim Y, Hwang JJ. Deep learning-based detection of apical lesions in panoramic radiographs: A comparative study. J Dent Sci. 2022;17(3):1324\u0026ndash;31. https://doi.org/10.1016/j.jds.2021.09.006\u003c/li\u003e\n \u003cli\u003eGaller KM, Weber M, Korkmaz Y, Widbiller M, Feuerer M. Inflammatory response mechanisms of the dentine\u0026ndash;pulp complex and the periapical tissues. Int J Mol Sci. 2021;22(3):1480. https://doi.org/10.3390/ijms22031480\u003c/li\u003e\n \u003cli\u003e\u0026Ccedil;elik B, Savaştaer EF, Kaya Hİ, \u0026Ccedil;elik ME. The role of deep learning for periapical lesion detection on panoramic radiographs. Dentomaxillofac Radiol. 2023;52(7):20230118. https://doi.org/10.1259/dmfr.20230118\u003c/li\u003e\n \u003cli\u003eScikit-learn. StratifiedShuffleSplit. Available from: https://scikit-learn.org/stable/modules/generated/sklearn.model_selection.StratifiedShuffleSplit.html [Accessed 27 July 2025].\u003c/li\u003e\n \u003cli\u003eTian Y, Ye Q, Doermann D. Yolov12: Attention-centric real-time object detectors. arXiv [Preprint]. 2025. arXiv:2502.12524. https://doi.org/10.48550/arXiv.2502.12524\u003c/li\u003e\n \u003cli\u003eUltralytics. YOLOv12 Documentation. Available from: https://docs.ultralytics.com/tr/models/yolo12/ [Accessed 27 July 2025].\u003c/li\u003e\n \u003cli\u003eLv H, Xiang Q, Xiao J. Object detection based on improved RT-DETR for human-robot collaboration manufacturing system. In: Proceedings of the Fourth International Conference on Mechanical, Electronics, and Electrical and Automation Control (METMS 2024). Bellingham, WA: SPIE; 2024. Vol. 13163, p. 2144\u0026ndash;8.\u003c/li\u003e\n \u003cli\u003eWei J, Shen W, Hu T, Liu Q, Yu J, Huang J. Dynamic-DETR: Dynamic perception for RT-DETR. In: Proceedings of the 2024 6th International Conference on Robotics and Computer Vision (ICRCV). New York, NY: IEEE; 2024. p. 28\u0026ndash;32.\u003c/li\u003e\n \u003cli\u003eKhanagar SB, Al-Ehaideb A, Maganur PC, Vishwanathaiah S, Patil S, Baeshen HA, et al. Developments, application, and performance of artificial intelligence in dentistry\u0026ndash;A systematic review. J Dent Sci. 2021;16(2):508\u0026ndash;22. https://doi.org/10.1016/j.jds.2020.06.018\u003c/li\u003e\n \u003cli\u003eBoztuna M, Firincioglulari M, Akkaya N, Orhan K. Segmentation of periapical lesions with automatic deep learning on panoramic radiographs: An artificial intelligence study. BMC Oral Health. 2024;24:1332. https://doi.org/10.1186/s12903-024-03925-6\u003c/li\u003e\n \u003cli\u003eHan T, Hou S, Gao C, Xu S, Pang J, Gu H, et al. EF-RT-DETR: An efficient focused real-time DETR model for pavement distress detection. J Real-Time Image Process. 2025;22:63.\u003c/li\u003e\n \u003cli\u003eHe W, Zhang Y, Xu T, An T, Liang Y, Zhang B. Object detection for medical image analysis: Insights from the RT-DETR model. In: Proceedings of the 2025 International Conference on Artificial Intelligence and Computational Intelligence. New York, NY: IEEE; 2025. p. 415\u0026ndash;20.\u003c/li\u003e\n \u003cli\u003eSchwendicke FA, Samek W, Krois J. Artificial intelligence in dentistry: Chances and challenges. J Dent Res. 2020;99(7):769\u0026ndash;74. https://doi.org/10.1177/0022034520915714\u003c/li\u003e\n \u003cli\u003eGaur M, Pal S, Chaudhuri R, Anto OB. Using artificial intelligence. In: Data Intelligence and Cognitive Informatics: Proceedings of ICDICI 2023. Cham, Switzerland: Springer; 2024. Vol. 327.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"AI in dentistry, deep learning, pediatric endodontics, furcation lesion, clinical decision-support, panoramic radiograph, RT-DETR, YOLOv12x","lastPublishedDoi":"10.21203/rs.3.rs-7322034/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7322034/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground: \u003c/strong\u003eFurcation lesions in primary molars are critical in pediatric dentistry as they often guide decisions between root canal treatment and extraction. This study introduces a novel deep learning-based clinical decision-support system that directly maps radiographic lesion characteristics to corresponding treatment recommendations—a first in the context of pediatric dental imaging.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods: \u003c/strong\u003eA total of 387 anonymized panoramic radiographs from children aged 3–13 was labeled into five distinct bone lesion categories. Three object detection models (YOLOv12x, RT-DETR-L, and RT-DETR-X) were trained and evaluated using stratified train-validation-test splits. Diagnostic performance was assessed using precision, recall, [email protected], and [email protected]–0.95. Additionally, qualitative accuracy was evaluated with expert-annotated samples.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults: \u003c/strong\u003eAmong the models, RT-DETR-X achieved the highest accuracy ([email protected] = 0.434). All models successfully identified lesion types and supported corresponding clinical decisions. The system reduced diagnostic ambiguity and showed promise in supporting clinicians with varying levels of experience.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusions: \u003c/strong\u003eThis study presents the first deep learning–based decision-support framework that links primary molar furcation lesion classification with treatment planning using panoramic radiographs. The proposed models have potential for standardizing diagnostic outcomes, especially in resource-limited settings and mobile clinical environments.\u003c/p\u003e","manuscriptTitle":"Panoramic Radiograph–based Deep Learning Models for Diagnosis and Clinical Decision Support of Furcation Lesions in Primary Molars","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-09-01 08:48:31","doi":"10.21203/rs.3.rs-7322034/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"f08f2430-8346-4780-b8bb-30452941f4a0","owner":[],"postedDate":"September 1st, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-09-12T05:23:40+00:00","versionOfRecord":[],"versionCreatedAt":"2025-09-01 08:48:31","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7322034","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7322034","identity":"rs-7322034","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}