Dental Panoramic Radiograph Analysis Using YOLO26 From Tooth Detection to Disease Diagnosis
| Authors | Khawaja Azfar Asif & Rafaqat Alam Khan |
| Year | 2026 |
| Field | Computer Vision |
| arXiv | 2604.16231 |
| Download | |
| Categories | cs.CV |
Abstract
Panoramic radiography is a fundamental diagnostic tool in dentistry, offering a comprehensive view of the entire dentition with minimal radiation exposure. However, manual interpretation is time-consuming and prone to errors, especially in high-volume clinical settings. This creates a pressing need for efficient automated solutions. This study presents the first application of YOLOv26 for automated tooth detection, FDI-based numbering, and dental disease segmentation in panoramic radiographs. The DENTEX dataset was preprocessed using Roboflow for format conversion and augmentation, yielding 1,082 images for tooth enumeration and 1,040 images for disease segmentation across four pathology classes. Five YOLOv26-seg variants were trained on Google Colab using transfer learning at a resolution of 800x800. Results demonstrate that the YOLOv26m-seg model achieved the best performance for tooth enumeration, with a precision of 0.976, recall of 0.970, and box mAP50 of 0.976. It outperformed the YOLOv8x baseline by 4.9% in precision and 3.3% in mAP50, while also enabling high-quality mask-level segmentation (mask mAP50 = 0.970). For disease segmentation, the YOLOv26l-seg model attained a box mAP50 of 0.591 and a mask mAP50 of 0.547. Impacted teeth showed the highest per-class average precision (0.943), indicating that visual distinctiveness influences detection performance more than annotation quantity. Overall, these findings demonstrate that YOLOv26-based models offer a robust and accurate framework for automated dental image analysis, with strong potential to enhance diagnostic efficiency and consistency in clinical practice.
Engineering Breakdown
Plain English
This paper applies YOLOv26, a state-of-the-art object detection model, to automate three critical tasks in dental radiography: identifying and locating individual teeth, assigning FDI numbering (the standard dental numbering system), and segmenting diseased regions. The researchers preprocessed the DENTEX dataset using Roboflow, creating 1,082 images for tooth detection and 1,040 images for disease classification across four pathology classes. They trained five variants of YOLOv26-seg using transfer learning on Google Colab, establishing the first baseline for automated analysis of panoramic dental X-rays. This addresses a real clinical bottleneck: manual interpretation is slow and error-prone when radiologists process high volumes of images daily.
Core Technical Contribution
The core novelty is the first application of YOLOv26 (a recent YOLO variant) specifically adapted for the dental domain with two complementary outputs: instance segmentation for precise tooth localization and multi-class disease segmentation. Rather than treating tooth detection and pathology separately, the authors use a unified segmentation architecture that outputs both tooth boundaries and disease masks simultaneously. The FDI-based numbering is derived from the spatial arrangement of detected teeth, leveraging the standard anatomical ordering in panoramic radiographs. This represents a departure from prior dental AI work that typically handled detection and classification as separate pipeline stages, reducing error propagation and system complexity.
How It Works
The pipeline begins with raw panoramic radiographs that are preprocessed via Roboflow, which handles format standardization and applies augmentation techniques (rotation, brightness adjustment, etc.) to expand the training set from limited raw images to 1,082 and 1,040 examples respectively. YOLOv26-seg is a real-time object detection model enhanced with segmentation masks; it processes input images at a fixed resolution and outputs bounding boxes with pixel-level segmentation masks for each tooth region. The model simultaneously predicts tooth class/location and disease class at each masked region, effectively doing multi-task learning in a single forward pass. Transfer learning is employed, meaning the base YOLO weights are pretrained on general object detection tasks and then fine-tuned on the dental dataset, reducing training time and data requirements compared to training from scratch. Five architectural variants were trained to explore trade-offs in speed versus accuracy, likely differing in backbone depth, anchor configurations, or loss weighting. Post-processing applies the FDI numbering convention (teeth are numbered 1–32 following anatomical position) to the detected tooth instances in left-to-right, upper-to-lower order.
Production Impact
In a clinical radiology workflow, this system would automatically flag teeth with detected pathologies (e.g., cavities, bone loss, impactions) and assign them FDI numbers, reducing radiologist read time from minutes per image to seconds of review-and-confirm. The unified segmentation approach is cheaper to deploy than cascaded pipelines (detection → classification → disease mapping) because it requires only one model inference per image instead of three, lowering GPU/CPU load and latency. Integration into existing PACS (Picture Archiving and Communication Systems) would be straightforward: the model outputs machine-readable annotations (bounding boxes, masks, FDI codes) that can populate radiology reports or trigger alerts for critical findings. Key trade-offs include: (1) YOLOv26 is compute-intensive, requiring GPU inference in most hospital settings; (2) the model must generalize across different X-ray machines and acquisition protocols, which may require periodic retraining on new data; (3) regulatory approval (510(k) clearance or equivalent) is necessary before clinical deployment, adding 6–18 months to productionization. The approach is most valuable in high-volume practices (>20 radiographs/day) where the time savings justify infrastructure investment.
Limitations and When Not to Use This
The dataset size (1,082 and 1,040 images) is modest by modern deep learning standards and may not capture the full distribution of rare pathologies, atypical dental anatomy, or images from newer/older X-ray equipment—this could lead to performance degradation in out-of-distribution clinical settings. The paper assumes panoramic radiographs are well-centered and of diagnostic quality; severely rotated, partially captured, or low-contrast images may cause the model to miss teeth or misassign FDI numbers, creating liability in clinical use. The four pathology classes are not explicitly named in the abstract, raising questions about disease coverage: does the model detect all clinically relevant conditions (implants, orthodontics, etc.), or only a narrow subset? Transfer learning on general object detection pretrained weights may introduce artifacts or biases not relevant to dental anatomy, and the abstract provides no ablation study comparing pre-training strategies. Finally, the paper does not address inter-observer variability or ground-truth annotation reliability—if human annotators disagreed on disease labels during dataset creation, the model learns from noisy supervision, potentially limiting its clinical utility.
Research Context
This work builds on the broader trend of applying deep learning to medical imaging, particularly the maturation of YOLO-family models (YOLOv5, YOLOv8, now YOLOv26) which have become the de facto standard for real-time segmentation in resource-constrained medical settings. The DENTEX dataset represents the community's first large-scale, multi-task dental radiograph benchmark, analogous to how COCO or ImageNet accelerated progress in general computer vision. Prior dental AI research relied on classical machine learning or older CNN architectures; this paper leapfrogs to state-of-the-art, establishing a new baseline that will likely spawn follow-up work on domain adaptation (handling different X-ray devices), semi-supervised learning (leveraging unlabeled clinical images), and regulatory compliance. The unification of detection and segmentation in a single model is aligned with current trends in medical AI toward end-to-end architectures that reduce error propagation and improve interpretability, setting a template for similar radiograph analysis tasks (maxillofacial, orthopedic, etc.).
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::