Research map / Detection and segmentation / Open-set and few-shot detection
OralXrays-9: Towards Hospital-Scale Panoramic X-ray Anomaly Detection via Personalized Multi-Object Query-Aware Mining
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 OralXrays-9, a 12,688-image panoramic dental X-ray dataset with 84,113 anomaly boxes across nine classes, and proposes MOQAM, a DETR-style detector with DI-RPN and CB-SCR modules for multi-scale, class-imbalanced oral anomaly detection.
Problem
Automated detection of multiple oral anomalies in panoramic dental X-rays is hard due to scale variation across lesions and severe class imbalance, and no large public multi-label dental dataset existed.
Approach
Adds an auxiliary one-to-many query head with scale-specific positive assignment (KL-divergence between effective receptive field Gaussian and GT box, plus size-dependent dynamic IoU thresholds) and a class-balanced hyperspherical contrastive loss with class prototypes to mine better positive queries.
Key results
- MOQAM reaches 73.5 AP on OralXrays-9 versus 71.0 for Salience-DETR and 69.8 for CFINet (best CNN).
- On SODA-D, MOQAM achieves 33.1 AP versus 32.2 for Co-DINO.
- On LVIS, MOQAM achieves 30.9 AP versus 29.8 for Co-DINO; ablation shows DI-RPN alone gives 72.3 AP and CB-SCR alone gives 72.0 AP on OralXrays-9.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add DI-RPN to the baseline detector (ablation) | MOQAM on OralXrays-9 · OralXrays-9 | AP: 70.9 -> 72.3 | Helped |
| Add CB-SCR to the baseline detector (ablation) | MOQAM on OralXrays-9 · OralXrays-9 | AP: 70.9 -> 72.0 | Helped |
| Add DI-RPN and CB-SCR together (ablation) | MOQAM on OralXrays-9 · OralXrays-9 | AP: 70.9 -> 73.5 | Helped |
| Vary base IoU threshold tb among 0.5, 0.55, and 0.6 (ablation) | MOQAM · OralXrays-9 | AP: 72.9; 73.5; 73.2 | Mixed |
| Vary growth exponent p among 0.1, 0.2, and 0.3 (ablation) | MOQAM · OralXrays-9 | AP: 73.3; 73.5; 73.1 | Mixed |
| Vary loss weight α among 0.3, 0.4, 0.5, 0.6, and 0.7 (ablation) | MOQAM · OralXrays-9 | AP: 72.9; 72.7; 73.4; 73.5; 73.2 | Mixed |
Limitations
- Compute hardware and training details beyond batch size and schedule are not reported in the main text.
- Ablation and hyperparameter analysis are reported only on OralXrays-9 and the reported gains are modest (about 1 to 3 AP).
Experimental setup
Datasets: OralXrays-9 SODA-D LVIS
Benchmarks: OralXrays-9 SODA-D LVIS
Metrics: AP AP50 AP75 APS APM APL APr APc APf
Baselines: Faster RCNN ATSS YOLOX Sparse RCNN CFINet DETR Deformable-DETR Conditional-DETR DAB-DETR DINO Co-DINO Salience-DETR
Builds on: Co-DINO DINO Salience-DETR Deformable-DETR
Closest papers by meaning
- Towards Better Dental AI: A Multimodal Benchmark and Instruction Dataset for Panoramic X-ray Analysis (NeurIPS 2025 · similarity 0.89)
- CariesXrays: Enhancing Caries Detection in Hospital-Scale Panoramic Dental X-rays via Feature Pyramid Contrastive Learning (AAAI 2024 · similarity 0.89)
- OralGPT-Omni: A Versatile Dental Multimodal Large Language Model (CVPR 2026 · similarity 0.87)
- OralGPT-Plus: Learning to Use Visual Tools via Reinforcement Learning for Panoramic X-ray Analysis (CVPR 2026 · similarity 0.87)
- NeBLa: Neural Beer-Lambert for 3D Reconstruction of Oral Structures from Panoramic Radiographs (AAAI 2024 · similarity 0.84)
- Segmenting Maxillofacial Structures in CBCT Volumes (CVPR 2025 · similarity 0.84)
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.