[Paper] DentalGPT: Incentivizing Multimodal Complex Reasoning in Dentistry

Source: arXiv - 2512.11558v1

Overview

DentalGPT is a domain‑specific multimodal large language model (MLLM) that can “see” dental images and reason about them like a specialist. By training on the largest publicly disclosed dental image‑text dataset (≈120 k paired samples) and fine‑tuning with reinforcement learning, the 7 B‑parameter model reaches or exceeds the performance of much larger general‑purpose MLLMs on dental diagnosis and visual‑question‑answering tasks.

Key Contributions

• Largest dental multimodal dataset – 120 k intra‑oral and panoramic images with detailed, diagnosis‑focused captions, released as a benchmark for the community.
• Two‑stage adaptation pipeline – (1) supervised fine‑tuning on the dental corpus to inject visual knowledge, followed by (2) reinforcement learning from human‑annotated reasoning traces to boost complex multimodal reasoning.
• Compact yet powerful model – A 7 B‑parameter transformer that outperforms many 30 B+‑parameter general MLLMs on dental VQA and disease‑classification benchmarks.
• Comprehensive evaluation suite – New intra‑oral and panoramic test sets plus dental subsets of existing medical VQA benchmarks, with metrics for classification accuracy, answer correctness, and reasoning fidelity.
• Open‑source release – Model weights, data, and training scripts are made publicly available to accelerate research and product development in oral health AI.

Methodology

1. Data Collection & Curation

   • Aggregated images from dental clinics, open‑source radiology archives, and educational repositories.
   • Each image was paired with a caption that explicitly names visual cues (e.g., “radiolucent lesion at the distal root of tooth #30”) and a short diagnostic rationale.
   • Quality control involved dental experts reviewing a random 5 % of the pairs for correctness and completeness.

2. Supervised Fine‑Tuning

   • Started from a pretrained vision‑language backbone (ViT‑Q‑former + LLaMA‑2‑7B).
   • Trained on the dental corpus using standard cross‑entropy loss to align image embeddings with the detailed captions.

3. Reinforcement Learning from Human Feedback (RLHF)

   • Collected “reasoning traces” where experts answered a VQA prompt step‑by‑step (e.g., “Identify the lesion → Compare with known patterns → Choose diagnosis”).
   • Used Proximal Policy Optimization (PPO) to reward model outputs that matched expert traces, encouraging chain‑of‑thought reasoning across modalities.

4. Inference Pipeline

   • At runtime, the model receives an image and a free‑form question.
   • The visual encoder extracts a dense representation, which the language decoder attends to while generating a step‑wise answer, optionally emitting a confidence score.

Results & Findings

• Parameter efficiency: Despite being ~4× smaller than competing models, DentalGPT closes >80 % of the performance gap to human experts.
• Fine‑grained visual understanding: Ablation studies show that the detailed captions improve detection of subtle pathologies (e.g., early caries, periapical radiolucencies) by >10 % relative to generic caption data.
• Reasoning boost: RLHF adds ~6–8 % absolute gain on VQA tasks, confirming that step‑wise supervision is critical for dental diagnostics.

Practical Implications

• Clinical decision support: Dental clinics can embed DentalGPT into imaging software to provide instant differential diagnoses, triage suggestions, or patient‑friendly explanations.
• Tele‑dentistry platforms: Automated pre‑screening of uploaded intra‑oral photos can flag urgent cases, reducing response latency for remote consultations.
• Education & training: Dental schools can use the model as an interactive tutor that explains radiographic findings and answers “why” questions, complementing human instructors.
• Regulatory‑ready pipelines: Because the model is compact, it fits on edge devices (e.g., dental chair‑side workstations) and can be audited more easily than massive black‑box models.
• Data‑centric AI workflow: The paper demonstrates a reproducible recipe—collect high‑quality domain data → supervised fine‑tune → RLHF—that can be replicated for other specialties (dermatology, ophthalmology, etc.).

Limitations & Future Work

• Dataset bias: The training set is dominated by images from a few geographic regions and equipment types, which may limit generalization to under‑represented populations.
• Explainability: While chain‑of‑thought outputs improve transparency, the underlying visual encoder remains a black box; future work could integrate attention visualizations or saliency maps.
• Regulatory validation: Clinical trials are needed to assess safety and efficacy before deployment in real patient care.
• Multimodal expansion: Current work focuses on static images; extending to video (e.g., intra‑oral scans) and 3‑D cone‑beam CT would broaden applicability.

DentalGPT shows that a well‑curated, domain‑specific multimodal dataset combined with staged fine‑tuning can produce a lightweight, high‑performing AI assistant for dentistry—opening the door for similar breakthroughs across healthcare.

Authors

• Zhenyang Cai
• Jiaming Zhang
• Junjie Zhao
• Ziyi Zeng
• Yanchao Li
• Jingyi Liang
• Junying Chen
• Yunjin Yang
• Jiajun You
• Shuzhi Deng
• Tongfei Wang
• Wanting Chen
• Chunxiu Hao
• Ruiqi Xie
• Zhenwei Wen
• Xiangyi Feng
• Zou Ting
• Jin Zou Lin
• Jianquan Li
• Guangjun Yu
• Liangyi Chen
• Junwen Wang
• Shan Jiang
• Benyou Wang

Paper Information

• arXiv ID: 2512.11558v1
• Categories: cs.CV, cs.AI, cs.CL
• Published: December 12, 2025
• PDF: Download PDF