TL;DR
PeriaTrace is a three-stage AI framework that automates the interpretation of periapical radiographs (PRs), generating tooth-level diagnoses linked directly to visual evidence. By combining deep learning for visual perception with large language models for reasoning, it produces both professional reports and patient-friendly explanations in seconds, significantly reducing clinician workload and improving patient comprehension.
What Problem Does This Paper Address?
Periapical radiographs (PRs) are essential for diagnosing dental issues like root canal infections and bone loss. However, interpreting these images is time-consuming and requires specialized expertise. Furthermore, standard dental reports are often filled with technical jargon (e.g., "periapical radiolucency") that leaves patients confused or anxious.
While many AI systems can detect dental anomalies, they often struggle with visual grounding—the ability to explicitly link a textual finding to a specific tooth or region in the image. Existing models might generate a correct diagnosis but fail to "point" to the evidence, making it difficult for clinicians to verify the output and for patients to visualize their condition.
Methods
The authors propose PeriaTrace, a modular three-stage pipeline designed for traceability and clarity:
- Stage 1: Multitask Visual Prior Extraction: Using a Segment Anything Model (SAM) with a ViT-B encoder, the system extracts "visual priors." These include tooth localization, Fédération Dentaire Internationale (FDI) numbering, semantic segmentation of structures (like pulp and bone), and lesion detection.
- Stage 2: Structured Diagnosis: These visual priors, along with the raw image, are fed into a Vision-Language Model (VLM)—specifically a LoRA-adapted Qwen3-VL-8B. A "Geometric Binding Engine" ensures that findings are strictly associated with the correct tooth-level evidence.
- Stage 3: Final Report Generation: A deterministic program takes the structured output from the VLM to generate three distinct components: a professional diagnostic report, a plain-language patient explanation, and actionable care recommendations.
Datasets
The study introduces MM-PRAD, the first large-scale multimodal dataset dedicated to PR analysis. It contains:
- 10,192 radiographs from 6,672 patients.
- Pixel-level segmentation masks and bounding boxes.
- Clinician-authored professional reports and corresponding patient-friendly versions.
For robustness, the model was also evaluated on two external clinical cohorts (TUMC and TMUSH) totaling over 2,600 additional radiographs.
Key Insights and Ideas
- Priors are Essential for Grounding: One of the most significant findings is that standard VLMs (even those trained on dental data like DentVLM) exhibit negligible grounding capability when given only raw images. Providing explicit visual priors (Stage 1) increased grounding F1 scores from near-zero to over 60%.
- The "Jargon Gap" is Real: A survey of 196 non-medical participants showed a massive preference (96.4%) for reports that include plain-language explanations. Patients were significantly better at identifying their own health issues and necessary follow-up actions when using the PeriaTrace format compared to standard professional reports.
- Traceability Builds Trust: By embedding machine-readable "evidence tags" in the text that link to specific image masks, the system allows clinicians to quickly verify the AI's reasoning, turning a "black box" into a transparent tool.
Key Results
- Accuracy: PeriaTrace achieved a macro Dice coefficient of 85.35% for segmentation and an F1 score of 94.37% for joint tooth localization and identification.
- Clinical Utility: In external evaluations, senior endodontists found that 84% to 91% of the generated reports were suitable for use with only minor or no revisions.
- Efficiency: The system generates a complete report in an average of 4.41 seconds. Clinicians estimated that manual drafting of an equivalent report takes 15–20 minutes, representing a massive speedup.
- Patient Impact: 84.7% of survey participants found the visual localization cues (pointing to the problem area) helpful for understanding their diagnosis.
Limitations and Future Work
- Geographical Bias: All data was collected from a single geographical region, which may affect generalizability to different imaging equipment or clinical practices elsewhere.
- Error Propagation: Because the system is staged, errors in the initial tooth localization or segmentation can cascade into incorrect diagnostic text.
- Rare Conditions: The authors note that future work should focus on "open-set recognition" to better handle rare or previously unseen dental abnormalities.
- Clinical Workflow: While the system is fast, its actual impact on long-term clinical outcomes and patient behavior remains to be established in prospective, real-world studies.
Summary updated