Source-linked AI summary
Automated Dental Caries Segmentation in Panoramic Radiographs Using Dual-Stage Deep Learning
Jihun Kim, Kyeonghun Kim, Jong-yeol Lee, Yeongseok Seo, Dohyun Chun
TL;DR
Early caries detection is challenging, and pixel-wise expert annotation limits the scalability of deep-learning approaches. This paper combines tooth localization with lesion segmentation and converts polygon labels into training masks. It reports strong segmentation metrics and reduced false positives compared with existing methods.
Problem
Early proximal lesions in posterior teeth are difficult to detect, while manual pixel-wise annotation requires expertise and time that constrain scalable supervised learning.
Source: §1 · p. 2Method
The framework uses Faster R-CNN to localize teeth and U-Net to segment caries, with polygon annotations transformed into binary masks for pixel-wise training.
Source: §1 · p. 2, §2.1 · p. 3, §2.2 · p. 4Results
IoU of 0.9013 and Dice coefficient of 0.9482 were reported, with significantly reduced false-positive rates compared with existing methods.
Source: §4 · p. 8Takeaways & Limitations
The findings support the potential for scalable automated caries detection that provides consistent diagnostic support and may reduce inter-clinician variability.
Source: §4 · p. 8
Abstract
from arXiv · showhide
Early detection of dental caries remains challenging due to limitations in traditional diagnostic methods, particularly for proximal lesions in posterior teeth. Deep learning models show promise for automated caries detection but face scalability constraints due to requirements for large volumes of expertly annotated training data. This study presents a dual-stage deep learning framework combining Faster R-CNN for tooth localization with U-Net for pixel-wise caries segmentation in panoramic radiographs. We developed a systematic transformation pipeline to convert large-scale polygon-annotated datasets into high-resolution binary segmentation masks, enabling pixel-wise supervised learning. The framework was trained using both expert-verified datasets and algorithmically processed labels from 3,000 panoramic images. Our approach achieved robust performance with an IoU of 0.9013, Dice coefficient of 0.9482, Recall of 0.9433, and Precision of 0.9774, demonstrating superior accuracy compared to existing methods while significantly reducing false-positive rates. The dual-stage framework effectively addresses data annotation bottlenecks in dental AI applications and demonstrates potential for scalable, automated caries detection systems that can improve diagnostic consistency and support clinical decision-making.
1 Introduction
Early caries detection is difficult, especially for proximal lesions in posterior teeth, while manual pixel-wise annotation constrains scalable deep learning. The study combines tooth localization and lesion segmentation with a pipeline for converting polygon labels, and reports strong segmentation performance with fewer false positives.
1 Introduction
- Proximal lesions in posterior teeth are difficult to detect, and manual pixel-wise annotation requires substantial expertise and time, limiting supervised-learning scalability.Traditional examinations also face low sensitivity for early lesions and variability across clinicians.Source: §1 · p. 2
- A transformation pipeline converts polygon-annotated datasets into high-resolution binary masks, combining expert-verified and algorithmically processed labels for pixel-wise training.The approach targets the annotation bottleneck while expanding the use of existing datasets.Source: §1 · p. 2
- IoU of 0.9013 and Dice coefficient of 0.9482 were achieved, with significantly reduced false-positive rates compared with existing methods.The authors describe the results as supporting scalable automated detection and more consistent diagnostic support.Source: §1 · p. 2
2 Material and methods
The methods use expert-annotated and polygon-annotated panoramic radiographs to train a two-stage system: Faster R-CNN localizes teeth, then U-Net segments lesions. Polygon annotations are converted into tooth-level mask samples, and model training and evaluation use specified optimizers, settings, and segmentation metrics.
2.1 Dataset
- The primary dataset contains 597 radiographs annotated by five dental experts, split into 497 training and 100 test images.Experts supplied tooth bounding boxes and pixel-wise binary lesion masks.Source: §2.1 · pp. 2–3
- The secondary dataset contains 3,000 panoramic radiographs with polygon annotations that require conversion into pixel-level masks.The images came from an AI Hub initiative involving 11 hospitals.Source: §2.1 · p. 3
- The preprocessing pipeline renders lesion polygons onto binary masks, crops paired images and masks to individual teeth, and visually checks alignment.The secondary dataset supplied 18,131 tooth-level samples for training after primary training on 597 expert-verified images.Source: §2.1 · p. 3
2.2 Network Architecture
- Faster R-CNN localizes teeth in panoramic radiographs, and U-Net segments caries within the detected tooth regions.The U-Net uses an EfficientNet-B0 encoder and produces binary lesion masks.Source: §2.2 · p. 4, Figure 1 · p. 4, Figure 3 · p. 5
- Faster R-CNN uses MobilenetV3 as its feature-extractor backbone, with anchors of varied sizes and aspect ratios for tooth detection.The architecture is illustrated with tooth detection and bounding-box cropping.Source: §2.2 · p. 4, Figure 2 · p. 4
2.3 Training Configuration and Evaluation
- Individual tooth masks are resized and placed at their recorded offsets and scales to assemble panoramic-sized segmentation outputs.The composite mask is initialized at the original radiograph dimensions.Source: §2.3 · p. 5
- Faster R-CNN and U-Net each trained for 100 epochs, with batch sizes of 4 and 32 and learning rates of 0.001 and 0.0001, respectively.Faster R-CNN used Adam; U-Net used AdamW with weight decay 0.001 and Dice Loss.Source: §2.3 · p. 5, Table 2 · p. 5
- Performance was evaluated using IoU, Dice coefficient, Recall, and Precision.Table 2 summarizes parameters applied to each model.Source: Table 2 · p. 5, §2.3 · p. 6
3 Results
The framework achieved strong test-set segmentation metrics and outperformed three established methods overall. Tooth-level and panoramic visualizations compare predicted masks with ground truth.
3.1 Segmentation Results
- 0.9013 IoU, 0.9482 Dice coefficient, 0.9433 Recall, and 0.9774 Precision were achieved on the test dataset.Training stabilized after 20 epochs, with final IoU plateauing at approximately 0.965.Source: §3.1 · p. 6
- Fig. 5 compares tooth-level predictions with ground truth across carious and healthy cases; predicted masks were integrated into panoramic radiographs.Healthy samples received empty masks, while predicted lesions were incorporated into full-arch visualizations.Source: §3.1 · p. 6, Figure 5 · p. 6
- Fig. 6 shows raw panoramic radiographs, manual references, and predictions, with green overlays locating lesions across the arch.The magnified view supports comparison of lesion locations and spatial coherence with dental anatomy.Source: §3.1 · p. 7, Figure 6 · p. 7
3.2 Comparative Analysis
- The proposed method showed superior overall performance across all metrics versus the Ensemble, Custom U-Net, and nnU-Net.The comparison describes a balance of precision and recall with spatial alignment between predictions and ground truth.Source: §3.2 · p. 7, Table 3 · p. 7
4 Conclusions
The conclusion presents a Faster R-CNN and U-Net framework alongside a polygon-label transformation pipeline that expands tooth-level training data. It reports strong segmentation performance and identifies dataset diversity and clinical integration as future work.
4 Conclusions
- Faster R-CNN localizes teeth before U-Net performs lesion segmentation in panoramic radiographs.The two-stage approach identifies tooth regions and segments lesions within those localized areas.Source: §4 · p. 8
- The pipeline converts polygon annotations into binary masks and expands 3,000 images into 18,131 tooth-level training samples.The training strategy integrates expert-verified and algorithmically processed labels to address annotation bottlenecks.Source: §4 · p. 8
- 0.9013 IoU, 0.9482 Dice coefficient, 0.9433 Recall, and 0.9774 Precision were reported, with superior accuracy and fewer false positives than existing methods.The authors describe potential support for consistent diagnostic assistance and treatment planning.Source: §4 · p. 8
- Future work will expand dataset diversity and validate clinical integration pathways for routine AI-assisted dental diagnostics.Source: §4 · p. 8