Deep Learning in Dental Caries Detection
In recent years, deep learning has revolutionized diagnostic approaches in medical imaging, including dentistry. This study builds upon prior efforts to enhance caries detection by addressing two key limitations: limited classification granularity and a narrow focus on local lesions. Leveraging advanced object detection architectures—YOLO-v8, YOLO-v9, and YOLO-NAS—the study introduces a novel framework capable of full-tooth instance detection with fine-grained classification based on the International Caries Detection and Assessment System (ICDAS). Furthermore, two correction strategies integrating background knowledge were developed to improve model stability and accuracy under real-world clinical conditions.
Tooth Instance-Based Detection Using YOLO Models
Traditional approaches often localize caries within a limited area of interest, which undermines comprehensive diagnosis. This research innovates by implementing a full-tooth instance detection method where every tooth in an intraoral image is first identified before classification. The use of YOLO-v8, YOLO-v9, and YOLO-NAS enabled precise bounding of tooth regions, setting the groundwork for accurate classification. This structured detection method allows AI to handle multi-tooth scenarios—something critical in clinical diagnosis—and contributes to robust and scalable dental assessment systems.
Fine-Grained Caries Classification with ICDAS Integration
A standout feature of the study is the incorporation of ICDAS, a standardized grading system for caries progression. This fine-grained classification moves beyond the binary "caries/no caries" output of most existing systems and allows detection of subtle differences across early, moderate, and severe caries stages. The detailed labeling of 8,754 images provides a rich dataset to train models in identifying nuanced dental conditions, thereby improving the relevance of AI-supported diagnostics in individualized patient care.
Postprocessing Corrections Using Domain Knowledge
To enhance model reliability under complex conditions, the study introduces two postprocessing correction techniques. One method involves weighted averaging of predicted category scores to reduce uncertainty, while the other adaptively modifies confidence levels based on the spatial arrangement of teeth. These corrections mimic expert decision-making by incorporating anatomical logic, leading to greater diagnostic consistency and improved performance metrics such as mAP, precision, and recall.
Performance Evaluation and Comparative Analysis
Extensive evaluation showed that the proposed model corrections significantly improved performance across all three YOLO models. Notably, YOLO-v8 reached a peak mAP of 72.9%, with a 4.7% improvement attributed to the correction techniques. The average increase in precision and recall (3.8% and 5.6%, respectively) underscores the clinical viability of the models. The slight drop in FPS from 83.1 to 78.1 represents a reasonable trade-off for enhanced diagnostic reliability, particularly in moderate caries cases where earlier systems struggled.
Clinical Implications and Future Research Directions
The success of this study has profound clinical implications. By offering fine-tuned, tooth-specific diagnostic insights, AI tools can now assist dentists in devising more personalized treatment plans. The results validate the integration of background knowledge into AI models as a pathway toward real-world applicability. This work lays the groundwork for future studies that might explore multimodal integration (e.g., combining radiographic and visual data), transfer learning for rare dental conditions, and real-time intraoral analysis through smart dental devices.
Technology Scientists Awards
Comments
Post a Comment