Research map / Vision-language and multimodal / Multimodal LLMs
Towards Better Dental AI: A Multimodal Benchmark and Instruction Dataset for Panoramic X-ray Analysis
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.
Introduces MMOral, a 20,563-image panoramic X-ray instruction dataset (1.3M instances) and MMOral-Bench, evaluates 64 LVLMs, and fine-tunes Qwen2.5-VL-7B into OralGPT, which gains 24.73% average score.
Problem
General medical LVLM benchmarks and instruction datasets omit dental panoramic X-ray interpretation, which has dense anatomy and subtle pathology cues.
Approach
Builds a dental instruction dataset from specialist-detected anatomical structures converted into grounding captions, LLM-generated reports (generation then GPT-4-turbo correction), and template and LLM-based QA/chat, then benchmarks LVLMs and SFTs Qwen2.5-VL-7B on it.
Key results
- GPT-4o achieves the best average-of-categories score of 41.45% on MMOral-Bench among 64 evaluated LVLMs, per the paper's abstract and table.
- OralGPT (Qwen2.5-VL-7B SFT on all MMOral data, one epoch) raises average score from 21.46% to 46.19%, a 24.73-point gain.
- Best medical-specific model HealthGPT-XL32 reaches 39.59% average, below general models such as Ovis2-34B (44.91%) and Kimi-VL-A3B-Thinking's 40.68%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| SFT Qwen2.5-VL-7B on MMOral-Report (ablation) | Qwen2.5-VL-7B · MMOral-Bench | average score: 10.35% | Helped |
| SFT Qwen2.5-VL-7B on MMOral-VQA (ablation) | Qwen2.5-VL-7B · MMOral-Bench | average score: 18.21% | Helped |
| Add MMOral-VQA to SFT with MMOral-Report (ablation) | Qwen2.5-VL-7B SFT on MMOral-Report · MMOral-Bench | average score: 21.46% to 44.53% | Helped |
| Add MMOral-Chat to SFT with MMOral-Report and MMOral-VQA (ablation) | Qwen2.5-VL-7B SFT on MMOral-Report and MMOral-VQA · MMOral-Bench | average score: 1.66% | Helped |
| Add MMOral-Chat to SFT with MMOral-Report and MMOral-VQA (ablation) | Qwen2.5-VL-7B SFT on MMOral-Report and MMOral-VQA · MMOral-Bench closed-ended and open-ended VQA | score | Mixed |
Surprising findings
- Medical-specific LVLMs (LLaVA-Med, HuatuoGPT-V, MedVLM-R1, MedDr, HealthGPT) do not outperform general-purpose models on dental panoramic X-rays.
- Some proprietary models (Qwen-Max-VL, Qwen-QVQ) refuse to answer dental questions citing safety or content policies.
Limitations
- Only panoramic X-rays are covered; other dental imaging modalities such as CBCT, periapical X-rays, and intraoral photos are absent.
- Open-ended evaluation depends on GPT-4-turbo judging, and dentist scores differ from it by about 1 point overall.
- Reports and QA pairs are LLM-generated from detector outputs, so errors from visual specialists and generators can propagate.
Experimental setup
Datasets: TED3 Hoang Viet Do apical periodontitis panoramic dataset MMOral
Benchmarks: MMOral-Bench
Metrics: accuracy LLM-judged soft score (0-1)
Baselines: GPT-4o GPT-4V Claude-3-7-Sonnet Gemini-2.5-Flash Gemini-2.0-Flash HealthGPT-XL32 MedDr Qwen2.5-VL-72B Ovis2-34B LLaVA-Med
Builds on: Qwen2.5-VL LLaMA-Factory VLMEvalKit MM-Vet
Closest papers by meaning
- OralGPT-Omni: A Versatile Dental Multimodal Large Language Model (CVPR 2026 · similarity 0.92)
- OralGPT-Plus: Learning to Use Visual Tools via Reinforcement Learning for Panoramic X-ray Analysis (CVPR 2026 · similarity 0.91)
- OralXrays-9: Towards Hospital-Scale Panoramic X-ray Anomaly Detection via Personalized Multi-Object Query-Aware Mining (CVPR 2025 · similarity 0.89)
- CariesXrays: Enhancing Caries Detection in Hospital-Scale Panoramic Dental X-rays via Feature Pyramid Contrastive Learning (AAAI 2024 · similarity 0.86)
- NeBLa: Neural Beer-Lambert for 3D Reconstruction of Oral Structures from Panoramic Radiographs (AAAI 2024 · similarity 0.86)
- Large-scale and Fine-grained Vision-language Pre-training for Enhanced CT Image Understanding (ICLR 2025 · similarity 0.86)
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.