OralGPT-Plus: Learning to Use Visual Tools via Reinforcement Learning for Panoramic X-ray Analysis

This paper introduces OralGPT-Plus, an agentic vision-language model that leverages a new dataset (DentalProbe), a reinspection-driven reinforcement learning framework, and a holistic benchmark (MMOral-X) to enable iterative, symmetry-aware, and clinically reliable diagnostic reasoning for panoramic dental radiographs.

The Gist

Imagine you are a detective trying to solve a mystery, but instead of looking at a crime scene, you are looking at a giant, flat X-ray of a human mouth. This is a Panoramic X-ray, and it's packed with tiny clues about cavities, bone loss, and hidden teeth.

For a long time, computer programs trying to read these X-rays were like novice detectives:

1. The Old Way (Object Detectors): They could point a finger and say, "There's a problem here," but they couldn't explain why or what kind of problem it was.
2. The Middle Way (Standard AI): They could write a paragraph describing the image, but they did it in one quick glance. If they missed a tiny crack in a tooth, they wouldn't go back to check. They just guessed and moved on.

OralGPT-Plus is the new Master Detective. It doesn't just look; it thinks, acts, and re-checks.

Here is how it works, broken down into simple concepts:

1. The "Smart Magnifying Glass" and the "Magic Mirror"

When a human dentist looks at an X-ray, they don't just stare at the whole picture. They zoom in on suspicious spots. But here's the secret trick: Dentists are symmetrical creatures. If a tooth on the left side looks weird, the dentist immediately looks at the matching tooth on the right side to see if it's normal.

• The Zoom Tool: OralGPT-Plus has a digital magnifying glass. If it sees a shadow, it zooms in to get a high-resolution look.
• The Mirror Tool (The Big Innovation): This is the paper's superpower. If the AI sees a weird spot on the left, it automatically flips the image to look at the right side. It asks, "Is the matching tooth on the other side healthy? If yes, then this spot on the left is probably a disease. If no, maybe it's just a weird angle."
  • Analogy: Imagine trying to see if a stain on your shirt is real. You look at the other side of the shirt. If the other side is clean, you know the stain is real. OralGPT-Plus does this instantly for every tooth.

2. Learning from a "Veteran Dentist" (Instruction Tuning)

You can't just tell a computer, "Be smart." You have to show it how.
The researchers created a special training dataset called DentalProbe. Think of this as a video game walkthrough or a cookbook written by expert dentists.

• Instead of just showing the AI the answer, they showed it the entire process: "First, look at the whole mouth. Second, zoom in on this tooth. Third, check the mirror image. Fourth, conclude."
• This taught the AI to mimic the step-by-step thinking of a real human doctor, rather than just guessing.

3. The "Coach" (Reinforcement Learning)

Once the AI learned the basics, the researchers needed to make it better at handling tricky cases. They used a Reinforcement Learning system, which is like having a strict but fair coach.

• The Problem: Sometimes the AI gets lazy. It might zoom in once, guess the answer, and stop. Or, it might zoom in 20 times on the same spot just to get points (a trick called "reward hacking").
• The Solution: The coach gives points based on a Rubric (a detailed checklist).
  • If the AI finds the disease correctly, it gets points.
  • If it misses a disease, it loses points.
  • Crucially: The coach only rewards the AI for re-checking (zooming or mirroring) if the first guess was already pretty good but incomplete. This stops the AI from wasting time looking at things that are already clear, and forces it to dig deeper only when necessary.

4. The New Test (MMOral-X)

To prove this new detective is better, the researchers built a new test called MMOral-X.

• Think of this as a final exam for dental AI.
• It has questions ranging from "Easy" (spot the obvious broken tooth) to "Hard" (find a tiny, hidden infection that looks just like normal bone).
• The results showed that OralGPT-Plus scored much higher than all the other AI models, including the expensive, famous ones from big tech companies.

The Big Picture

This paper is about changing how AI diagnoses dental problems.

• Before: AI was like a student who reads a book once and takes a test.
• Now (OralGPT-Plus): AI is like a senior doctor. It has a magnifying glass, it uses a mirror to compare sides, it follows a checklist, and it knows when to double-check its work.

By teaching the AI to act (use tools) and reflect (re-examine), the researchers have created a system that is much safer, more accurate, and much closer to how a human dentist actually works. It's not just about seeing the image; it's about understanding the story behind the image.

Technical Summary

1. Problem Statement

Panoramic dental radiographs (OPGs) are critical for oral diagnosis but present unique challenges for current Vision-Language Models (VLMs):

• Fine-Grained Complexity: Diagnosing dental issues requires understanding subtle spatial relationships, bilateral symmetry, and multi-step verification.
• Limitations of Static Paradigms: Traditional object detectors provide only bounding boxes without clinical reasoning. Standard VLMs operate under a single-pass paradigm, generating a static description in one forward pass. This prevents them from revisiting ambiguous regions, magnifying low-contrast lesions, or performing contralateral (left-right) comparisons, which are standard clinical practices for dentists.
• Lack of Clinical Alignment: Existing models often fail to emulate the iterative "thought-action-observation" workflow of human dentists, leading to unreliable or hallucinated diagnoses in complex cases.

2. Methodology

The authors propose OralGPT-Plus, an agentic VLM designed to perform iterative, symmetry-aware diagnostic reasoning. The framework consists of three core components:

A. DentalProbe Dataset & Instruction Tuning

To teach the model dentist-like behavior, the authors constructed DentalProbe, a dataset of 5,000 panoramic images with expert-curated diagnostic trajectories.

• Trajectory Construction: The data generation pipeline uses a multi-agent system to simulate clinical workflows. It starts with global inspection, identifies suspicious regions, and then iteratively applies tools to zoom in or compare symmetric regions.
• Dentist-like Instruction Tuning: The model is fine-tuned using Supervised Fine-Tuning (SFT) on these trajectories to learn when and how to use diagnostic tools.

B. Visual Tool Design

OralGPT-Plus is equipped with two specific tools that mimic clinical habits:

1. Zoom-In: Magnifies suspicious regions for detailed inspection of subtle lesions (e.g., early caries, apical periodontitis).
2. Mirror-In (Novelty): A symmetry-aware tool that retrieves the horizontally mirrored counterpart of a selected region across the midline. This allows the model to perform contralateral comparison, verifying if a shadow or irregularity is a true lesion or an artifact by comparing it to the healthy opposite side.

C. Reinspection-Driven Reinforcement Learning (RL)

To move beyond static instruction tuning, the authors developed a specialized RL framework to stabilize long-horizon reasoning and encourage meaningful re-examination.

• Rubric-Based Reward ($R_{rubrics}$): Instead of binary (0/1) rewards, a few-shot LLM evaluator assigns a continuous score (0.0–1.0) based on clinical accuracy, completeness, and the absence of major errors. This addresses reward sparsity in multi-lesion scenarios.
• Conditioned Diagnostic-Driven Reward ($R_{diag}$): To prevent "reward hacking" (where the model randomly uses tools to get points), exploration is gated. The curiosity bonus is only activated if the initial diagnosis is already reliable (high rubric score) but potentially incomplete. This ensures tools are used only when clinically justified.
• Hybrid Reward System: Combines rubric scores, format adherence, and the conditioned curiosity bonus to guide the agent toward stable, clinically coherent reasoning.

3. Key Contributions

1. OralGPT-Plus Model: The first agentic VLM for panoramic dental analysis that performs iterative diagnostic reasoning using "Zoom-In" and "Mirror-In" tools.
2. DentalProbe Dataset: A 5,000-image dataset with 8,000+ rounds of expert-guided diagnostic trajectories, enabling instruction tuning for structured tool usage and symmetry-aware reasoning.
3. MMOral-X Benchmark: The first holistic benchmark for panoramic diagnosis, containing 300 open-ended Q&A pairs across three difficulty levels (Simple, Moderate, Complex) with region-level annotations.
4. Reinspection-Driven RL Framework: A novel training strategy using continuous rubric-based rewards and confidence-gated exploration to stabilize multi-step reasoning and prevent unproductive tool usage.

4. Experimental Results

The model was evaluated on MMOral-X and the existing MMOral-OPG benchmark, outperforming strong baselines including proprietary models (GPT-5, Gemini, Claude), medical-specific VLMs (MedDr, HuatuoGPT-V), and open-source models.

• Performance: OralGPT-Plus (7B) achieved the highest overall score on MMOral-X (45.35), significantly surpassing the next best baseline (GPT-5 at 42.34) and medical-specific models (e.g., MedDr at 26.20).
• Ablation Studies:
  • Instruction Tuning: Essential for activating tool usage; models without it failed to use tools effectively even after RL.
  • Mirror-In Tool: Removing this tool caused a significant drop in accuracy (e.g., Complex cases dropped from 24.96 to 14.30), proving the value of symmetry-aware verification.
  • RL Rewards: Removing the conditioned diagnostic reward led to "reward hacking," where the model used tools indiscriminately, degrading performance.
• Scalability: The 7B model showed substantial gains from RL, while the 3B model struggled to sustain multi-turn reasoning, indicating that model capacity is crucial for mastering agentic workflows.

5. Significance

• Paradigm Shift: The work transitions dental AI from static perception (detecting boxes or describing images once) to interactive, iterative reasoning that mirrors human clinical workflows.
• Clinical Reliability: By incorporating "Mirror-In" and reinspection mechanisms, the model reduces false positives caused by anatomical asymmetries or artifacts, a critical step toward clinical deployment.
• Generalizability: The framework of agentic reasoning with domain-specific tools (symmetry, zooming) and specialized RL rewards offers a blueprint for applying VLMs to other medical imaging domains requiring fine-grained, multi-step analysis (e.g., mammography, CT scans).

In conclusion, OralGPT-Plus demonstrates that equipping VLMs with clinical reasoning patterns, specific diagnostic tools, and reinforcement learning tailored to medical constraints significantly enhances diagnostic accuracy and reliability in panoramic radiography.