Research map / Detection and segmentation / CNN features
CariesXrays: Enhancing Caries Detection in Hospital-Scale Panoramic Dental X-rays via Feature Pyramid Contrastive Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Releases CariesXrays, a 6,000-image panoramic dental X-ray caries detection benchmark, and proposes FPCL, a Faster R-CNN-based detector with a dual-directional feature pyramid and proposal-prototype contrastive regularization.
Problem
Automated detection of dental caries in panoramic dental X-rays is hard because lesions are small and have diverse appearances, and no public standardized dataset existed.
Approach
Combines a dual-directional feature pyramid network (D2D-FPN) with proposal-to-prototype contrastive regularization (P2P-CRL) on top of Faster R-CNN, with a momentum-based dynamic weighting of the contrastive loss, and a new hospital-scale caries dataset.
Key results
- FPCL achieves 48.2 AP on CariesXrays, versus 40.5 for YOLOx (best CNN baseline shown) and 42.2 for Conditional-DETR.
- Faster R-CNN baseline reaches 39.9 AP; adding FPN, CRL and DOT in the ablation brings the full model to 48.2 AP.
- Removing D2D-FPN, P2P-CRL, or dynamic optimization each lowers AP to 47.8 or 47.3 (from 48.2).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FPCL instead of Conditional-DETR | CariesXrays caries detection · CariesXrays | AP: 42.2 -> 48.2 | Helped |
| Use FPCL instead of YOLOx | CariesXrays caries detection · CariesXrays | AP50: 81.3 -> 84.1 | Helped |
| Add D2D-FPN (ablation) | FPCL · CariesXrays | AP: 47.3 -> 48.2 | Helped |
| Add P2P-CRL (ablation) | FPCL · CariesXrays | AP: 47.8 -> 48.2 | Helped |
| Add dynamic optimization training strategy (ablation) | FPCL · CariesXrays | AP: 47.8 -> 48.2 | Helped |
Limitations
- Ablation gains from individual components are small (about 0.4 to 0.9 AP).
- Deeper FPN configurations are resource-intensive and did not consistently improve results beyond 3 layers.
- Evaluation reported on a single dataset from clinical sources.
Experimental setup
Datasets: CariesXrays
Benchmarks: CariesXrays
Metrics: AP AP50 AP75 APS APM APL
Baselines: SSD RetinaNet DETR EfficientDet FCOS YOLOv7 YOLOv8 YOLOx Conditional-DETR Faster R-CNN
Builds on: Faster R-CNN EfficientDet (BiFPN) SimCLR MoCo DetCo
Closest papers by meaning
- OralXrays-9: Towards Hospital-Scale Panoramic X-ray Anomaly Detection via Personalized Multi-Object Query-Aware Mining (CVPR 2025 · similarity 0.89)
- Towards Better Dental AI: A Multimodal Benchmark and Instruction Dataset for Panoramic X-ray Analysis (NeurIPS 2025 · similarity 0.86)
- OralGPT-Plus: Learning to Use Visual Tools via Reinforcement Learning for Panoramic X-ray Analysis (CVPR 2026 · similarity 0.85)
- OralGPT-Omni: A Versatile Dental Multimodal Large Language Model (CVPR 2026 · similarity 0.84)
- Novel Class Discovery in Chest X-rays via Paired Images and Text (AAAI 2024 · similarity 0.82)
- Enhanced Contrastive Learning with Multi-view Longitudinal Data for Chest X-ray Report Generation (CVPR 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.