cs.CVMar 6, 2026
SaveOralGPT-Plus: Learning to Use Visual Tools via Reinforcement Learning for Panoramic X-ray Analysis
Organizations: The Hong Kong University of Science and Technology (GZ) · School of Computer Science, Peking University · Faculty of Dentistry, The University of Hong Kong · Shanghai Jiao Tong University · Institute of Automation, Chinese Academy of Sciences · College of Computer Science and Software Engineering, Shenzhen University
Abstract
Panoramic dental radiographs require fine-grained spatial reasoning, bilateral symmetry understanding, and multi-step diagnostic verification, yet existing vision-language models operate under a static single-pass paradigm that limits their clinical reliability. In this paper, we introduce OralGPT-Plus, an agentic vision-language model designed to perform iterative and symmetry-aware diagnostic reasoning for panoramic dental radiograph analysis. To support this paradigm, we construct DentalProbe, a five-thousand-image dataset with expert-curated diagnostic trajectories that provide structured supervision for localized inspection and contralateral comparison. We further develop a Reinspection-driven reinforcement learning framework that encourages clinically meaningful re-examination and stabilizes long-horizon reasoning with rubric-based reward and conditioned diagnostic-driven reward. In parallel, we present MMOral-X, the first benchmark for holistic panoramic diagnosis, containing 300 open-ended questions and region-level annotations across multiple difficulty levels. OralGPT-Plus demonstrates consistent and reliable improvements over strong baselines on MMOral-X and established panoramic benchmarks, indicating the effectiveness of interactive and symmetry-informed reasoning. Our work highlights the value of agentic modeling for dental imaging and provides a foundation for future research in clinically aligned panoramic radiograph analysis.
Figures & tables
Figure 1: Paradigm shift in panoramic dental radiograph analysis. Traditional detectors provide only category-level boxes, while VLMs generate coarse descriptions without structured reasoning. In contrast, OralGPT-Plus performs dentist-like diagnostic reasoning by invoking zoom-in and mirror-in tools to identify subtle findings and produce clinically coherent interpretations.
Figure 2: Curation process of DentalProbe dataset with expert trajectory for dentist-like instruction tuning.
Figure 3: Average score under dentist sampling evaluation on curated DentalProbe dataset.
Figure 4: Our training pipeline from dentist-like instruction tuning to re-inspection reinforcement learning.
Figure 5: Diagnostic trajectory generated by OralGPT-Plus on a sample of MMOral-X . The model produce a final panoramic-level diagnosis and got a score 0.9.
| Model | Params | MMOral-X | MMOral-OPG Open-ended VQA | ||||||||
| Simple | Moderate | Complex | Teeth | Patho | His | Jaw | Summ | Report | Overall | ||
| Proprietary VLMs (API) | |||||||||||
| GPT-5 [ 32 ] | N/A | 32.90 | 8.30 | 9.30 | 39.84 | 29.42 | 44.08 | 77.38 | 40.06 | 28.16 | 42.34 |
| Gemini-2.5-Flash [ 41 ] | N/A | 25.56 | 9.70 | 8.70 | 28.21 | 24.64 | 31.95 | 47.74 | 13.05 | 16.73 | 27.89 |
| Doubao-seed-1-6 a [ 9 ] | N/A | 20.10 | 5.80 | 5.20 | 27.19 | 18.18 | 41.72 | 59.44 | 22.38 | 40.20 | 34.20 |
| Claude-sonnet-4-5 b [ 1 ] | N/A | 27.72 | 8.14 | 9.78 | 28.29 | 20.45 | 48.83 | 71.56 | 26.67 | 37.70 | 37.68 |
Table 1: Model performance on MMOral-X and open-ended VQA for MMOral-OPG . Results on MMOral-X are averaged over five runs, whereas MMOral-OPG is evaluated once. In each column, the best score is bold and the second best is underlined .
Figure 6: Distribution across different task of different models and tool invocation comparison. (a) The performance of the MMOral-X benchmark. (b) Performance comparison on MMOral-Omni benchmark. (c) Performance comparison on MMOral-OPG benchmark.
| Model | MMOral-X | ||
| Simple | Moderate | Complex | |
| wo Instruction tuning | 8.02 | 4.64 | 4.98 |
| wo and | 20.82 | 9.24 | 9.82 |
| wo | 21.48 | 12.08 | 13.42 |
| wo | 24.62 | 14.14 | 15.96 |
| wo “Mirror-In” tool | 34.68 | 14.26 | 14.30 |
Table 2: Ablation studies.
Figure 7: Training dynamics across ablation models.
| Some categories of radiograph findings and abnormal conditions | ||
| Apical Periodontitis | Periapical Abscess | Periapical Granuloma |
| Apical Sclerosing Osteitis | Apical Rarefying and Sclerosing Osteitis | Focal Osteomyelitis |
| Pericoronitis | Periodontal Disease / Bone Loss | Endo-Perio Combination Lesion |
| Implantitis / Peri-implantitis | Mucous Retention Pseudocyst | Osteoarthritis of the Condyle |
| Tonsilloliths | Calcified Carotid Atheromatous Plaque | Antrolith |
| Foreign Body Reaction | Oroantral Communication | Trauma from Occlusion |
Table 3: Categories of radiograph findings and abnormal conditions.
| Dataset | Images Used | Annotation Scope | Region |
| Tufts [ 34 ] | 380 | Tooth & lesion labels; 5-level taxonomy | USA |
| Children [ 51 ] | 69 | Pediatric segmentation & disease labels | China |
| DENTEX [ 10 ] | 678 | Quadrant, enumeration, diagnosis (4 types) | Switzerland/Turkey/USA |
| Diagnostics TEAM [ 31 ] | 1435 | Detection of 12 clinical conditions | Romania |
| Total | 2562 | Fine-grained abnormality annotations | Multi-region |
Table 4: Summary of Integrated Public Dental Datasets
| Category | Radiographic Description |
| Abnormal tooth development | On panoramic radiographs, developmental abnormalities may involve tooth number, size, shape, or internal structure. Common findings include supernumerary or missing teeth, macrodontia, microdontia, dilaceration, dens in dente, ankylosis indicated by loss of periodontal ligament space, impacted or unerupted teeth, and enlarged pulp chambers such as taurodontism. |
| Caries | Caries typically presents as a radiolucent area within the enamel or dentin. Shallow lesions affect the enamel or outer dentin, moderate lesions reach the mid-dentin, and deep lesions extend toward the pulp. Early lesions may be subtle due to overlapping anatomy and reduced resolution in panoramic imaging. |
| Deep pits and fissures | Deep occlusal pits and fissures are often poorly visualized on panoramic radiographs because anatomical overlap hides early decay. Only advanced lesions penetrating dentin may appear as radiolucent defects, while early changes remain masked. |
| Periapical periodontitis | Appears as a radiolucent area surrounding the root apex or as widening of the periodontal ligament space. Early lesions may only present slight thickening or loss of lamina dura and can be difficult to detect due to superimposed anatomical structures. |
| Pulpitis | Chronic pulpitis may show widening of the periodontal ligament space or discontinuity of the lamina dura. Advanced pulpal inflammation may lead to periapical radiolucency or areas of increased bone density (condensing osteitis). |
| Impacted tooth | An impacted tooth appears as an unerupted or malpositioned tooth within the jawbone. Its outline may be partially obscured or distorted by overlapping structures, resulting in ghost images, blurry contours, or uncertain boundaries. |
Table 5: Categories and their radiographic descriptions retrieved from wiki and google webs.
| Category | Finding Type | Appearing Round(s) |
| Obvious findings | Prosthetic restoration | Round 1 |
| Orthodontic device | Round 1 | |
| Surgical device | Round 1 | |
| Implant | Round 1 | |
| Impacted tooth | Round 1 | |
| Bone-based findings | Bone resorption | Round 2+ |
Table 6: Three categories of radiographic findings used in our constructed multi-turn trajectories of DentalProbe dataset. Obvious findings appear only in Round 1, bone-based findings and subtle findings appear in Round 2 and onward, Due to the large number of subtle findings, we display only a subset of this type.
| Dataset Source | Task Type | Category Space | Categories | Training Set | mAP | AP50 |
| [ 10 , 34 , 6 ] | Object Detection | 1 to 32 tooth numbering following the FDI tooth numbering system | 32 | 2798 | 66.1 | 98.9 |
Table 7: The details of the teeth ID detection model we use.
Figure 8: Rounds distribution after RL of OralGPT-Plus 7B and 3B.
Figure 9: Evolution of key training dynamics throughout reinspection-driven reinforcement learning.
| Model | Mean | StdDev | CV % |
| GPT-5 [ 32 ] | 32.88 | 0.18 | 0.54 |
| Gemini-2.5-Flash [ 41 ] | 25.60 | 0.24 | 0.96 |
| GLM-4.5v [ 21 ] | 15.70 | 0.33 | 2.16 |
| Qwen2.5-VL-7B-Instruct [ 2 ] | 7.20 | 0.34 | 4.71 |
Table 8: Stability verification of using LLMs as judges: Standard deviation and coefficient of variation (CV) are reported across four VLMs from five repeated evaluations.
| Hyperparameter | Value |
| Epoch | 3 |
| LoRA Rank | 8 |
| LoRA | 16 |
| LoRA Dropout | 0.1 |
| LoRA Target | all |
| GPU | 4 NVIDIA A800 |
Table 9: Hyperparameters for training Qwen2.5-VL (7B and 3B) models in dentist-like instruction tuning.
| Hyperparameter | Value |
| Data & Prompt Settings | |
| System Prompt | tool_crop_mirror |
| Train Batch Size | 8 |
| Max Prompt Length | 8192 |
| Max Response Length | 8192 |
| Image Key | images |
Table 10: Key hyperparameters for resinspection-driven RL training in OralGPT-Plus.
Figure 10: Case of clustered proposal.
Figure 11: Case of clustered proposal.
Figure 12: Case of clustered proposal.
Figure 13: Case of clustered proposal.
Figure 14: Constructed trajectories for dentist-like instruction tuning.
Figure 15: Constructed trajectories for dentist-like instruction tuning.
Figure 16: Constructed trajectories for dentist-like instruction tuning.
Figure 17: Case of MMOral-X.
Figure 18: Case of MMOral-X.
Figure 19: System prompt for rubrics-based reward.
Figure 20: Query Prompt with Few-Shot Clinical Scoring Examples for rubrics-based reward.
Figure 21: System Prompt for Answer Agent to do Radiographic Analysis.
Figure 22: Query Prompt for Answer Agent to do Cropped Region Analysis
Figure 23: Query Prompt for Answer Agent to do Mirror Radiographic Analysis
Figure 24: Query Prompt for Judge Agent to do Output Verification for Tool Decision.