- The paper introduces a unified deep learning framework that achieves state-of-the-art performance in tooth enumeration and disease segmentation using YOLO26 on the DENTEX dataset.
- It details an end-to-end pipeline that converts DENTEX images to a YOLO26-compatible format without augmentation, leveraging advanced techniques like NMS-free detection, ProgLoss, and STAL.
- The study demonstrates significant improvements in precision and mAP metrics over previous models, setting new benchmarks for automated dental radiograph analysis.
Automated Analysis of Dental Panoramic Radiographs with YOLO26: Enumeration and Disease Segmentation on DENTEX
Introduction
The study proposes a unified deep learning framework for automated analysis of dental panoramic radiographs, exploiting the architectural advances introduced by YOLO26 for object detection and instance segmentation. Specifically, the work targets two core tasks in dental radiograph interpretation: (1) tooth localization, enumeration, and FDI-based numbering; (2) multi-class disease segmentation across four clinical pathologies. The pipeline is evaluated on the DENTEX benchmark, which remains the most comprehensive dataset with hierarchical annotations for quadrant identification, tooth enumeration, and disease classification in panoramic X-rays.
Methodology
The pipeline leverages an end-to-end automated preprocessing and training regimen. DENTEX images and annotations, natively in COCO JSON format, are converted to YOLO26-compatible format using Roboflow, ensuring consistent orientation and aspect ratio preservation (images standardized to 2048×1010, training at 800×800 due to GPU constraints). No augmentation is employed to preserve clinical presentation fidelity.
For tooth enumeration, images are labeled into four categories (canine, incisors, molar, premolar), following FDI notation. For pathology segmentation, four disease classes are used: caries, deep caries, impacted teeth, and periapical lesions. Five YOLO26-seg model scales are evaluated on each task, all trained with transfer learning on AdamW for 200 epochs without early stopping.
YOLO26 introduces NMS-free detection, ProgLoss for imbalance mitigation, STAL for small-object sensitivity, and the MuSGD optimizer for stable convergence with diverse radiograph statistics. These innovations are highly relevant given the unique challenges of dental imaging (e.g., small, overlapping structures, severe class imbalance, and subtle disease manifestations).
Results
Tooth Enumeration
YOLO26m-seg achieved the highest performance (precision = 0.976, recall = 0.970, box mAP50 = 0.976, mask mAP50 = 0.970), outperforming both lighter (YOLO26n) and heavier (YOLO26x) variants, demonstrating that moderate-capacity models represent a hardware/accuracy optimum under DENTEX data constraints. Notably, YOLO26m-seg surpasses previous best results from YOLOv8x by 4.9% in precision and 3.3% in mAP50.
Per-class analysis reveals highest AP for incisors (0.990), with molars, canines, and premolars following. Importantly, AP was driven by visual discriminability rather than annotation volume—premolars (despite a larger annotation count than canines) underperformed due to morphological similarity with molars.
Confusion matrices and PR curves further confirm high reliability and minimal inter-class confusion. All model scales achieve inference latency well under 45 ms, indicating clinical suitability for real-time applications even in resource-constrained deployment scenarios.
Disease Segmentation
YOLO26l-seg yielded the best disease segmentation performance (box mAP50 = 0.591, mask mAP50 = 0.547). Disease class performance varied dramatically: impacted teeth achieved AP = 0.943, while caries (AP = 0.463), deep caries (AP = 0.574), and periapical lesions (AP = 0.384) performed significantly worse. Crucially, visual distinctiveness—not annotation quantity—was the dominant factor: impacted teeth, although fewer in number, were radiographically unambiguous, while caries and periapical lesions (despite higher prevalence) remain challenging due to radiographic subtlety and high inter-class ambiguity.
Confusion matrix analysis revealed that less distinctive classes are often misclassified as background, limiting sensitivity for subtle pathologies. Training and validation loss curves confirm greater optimization instability on the segmentation task, primarily due to class imbalance and scarce positive examples for certain diseases. The presented box PR analysis visualizes the detection tradeoff and exposes the instability in decision boundary formation for overlapping or radiographically ambiguous diseases.
Comparative Evaluation
Against established benchmarks, the YOLO26m-seg model robustly outperformed Faster R-CNN and YOLOv8x on tooth enumeration, setting new state-of-the-art figures on DENTEX, and introduced mask-level segmentation for further clinical utility (absent in prior works). YOLO26l-seg is the first published model to address multi-class disease segmentation on the DENTEX benchmark, establishing baseline performance metrics and analytical methodology for future studies.
Implications and Future Directions
Practically, the pipeline offers a rapid, reproducible, and highly accurate tool for automatic panoramic radiograph analysis, with immediate implications for workload reduction, standardization of tooth numbering, and automated disease flagging, especially in high-throughput screening or regions with limited radiology expertise. From a theoretical perspective, the results establish that annotation volume alone does not guarantee robust detection for visually ambiguous pathologies, highlighting the critical need for targeted annotation strategies and, possibly, the integration of uncertainty quantification.
Key limitations include class imbalance (notably the rarity of periapical lesions), training at sub-maximal resolution due to compute constraints, and validation solely on DENTEX (without cross-device/institutional externalization). Future development will prioritize data balancing, high-resolution training, external validation, and integration into comprehensive clinical decision support workflows, validated via prospective practitioner studies.
Conclusion
The study delivers a systematic, reproducible pipeline for joint tooth enumeration and disease segmentation in dental panoramic radiographs, leveraging recent innovations in YOLO26 for practical clinical utility. Substantial improvement over prior state-of-the-art on the DENTEX benchmark is demonstrated, along with an analytical framework for class-level performance interpretation. By highlighting the primacy of visual distinctiveness in detection accuracy, the work delineates clear directions for ongoing research in dental AI, including dataset curation, model development for rare or ambiguous pathologies, and clinical deployment studies.
Reference:
"Dental Panoramic Radiograph Analysis Using YOLO26 From Tooth Detection to Disease Diagnosis" (2604.16231)