Skip to main content
Back to timeline
arXivSource publication:

PeriaTrace pins each periapical-radiograph finding to a named tooth, and outside endodontists kept 91.16% and 84.52% of reports across two cohorts

Synopsis

Periapical radiographs are the workhorse of endodontic imaging, yet reading them and writing reports is slow, conventional terminology confuses patients, and existing dental report generators rarely link a finding to a specific tooth and its radiographic evidence. PeriaTrace runs in three stages: it extracts tooth, caries and segmentation priors, has a vision-language model reason over them for tooth-level diagnosis, then validates and renders professional reports, patient explanations and care advice with evidence tags. On internal data, joint tooth localization and FDI identification reached 94.37% F1 and grounding-tag F1 was 77.93%; in two external cohorts, two senior endodontists judged 91.16% and 84.52% of reports retainable with revision rather than rewritten.

Source-provided article image: PeriaTrace: Tooth-Level Grounding Diagnosis and Patient-Facing Explanations from Periapical Radiographs
arXiv

Interpretation

Automated periapical-radiograph reporting can now attach each diagnostic statement to a specific tooth and image region instead of producing one global description. Prior dental vision-language models showed almost no such correspondence on raw radiographs, with grounding-tag F1 of only 0 to 5.70% across six baselines. Under one shared reporting template, giving those six baselines the same Stage 1 visual priors raised their grounding F1 to 57.96–62.96%, while PeriaTrace reached 77.93%, 14.97 percentage points above the strongest baseline.

The system first performs multitask visual analysis, then lets a language model reason over structured evidence, and finally applies rule-based validation and templates to render the report. Stage 1 uses a shared SAM ViT-B encoder with four task heads for tooth localization and FDI numbering, caries detection, nine-class foreground segmentation, and a quality-control classification of whether absolute FDI numbering is reliable. Stage 2 associates caries and segmentation components with teeth by overlap, containment and centroid distance, and feeds the radiograph, an annotation overlay and these priors to a LoRA-adapted Qwen3-VL-8B; Stage 3 removes invalid tooth identifiers and unsupported diagnostic codes before rendering tagged reports.

The quality-control head decides whether a report uses absolute FDI numbers or relative identifiers, so a wrong number does not propagate into later diagnoses. When absolute numbering is unreliable, the system switches to relative identifiers ordered left to right across visible teeth rather than forcing a possibly incorrect FDI number. That quality-control head scored 74.18% F1 on the internal test set, with 60.36% sensitivity and 99.53% specificity.

Perspective

Evaluation covers 10,192 periapical radiographs in the internal MM-PRAD cohort at the First Affiliated Hospital of Nankai University and external cohorts of 1,742 and 943 radiographs at two Tianjin hospitals; the external cohorts contain only clinician ratings of generated outputs, with no expert segmentation masks or reference reports. Patient-side findings come from a web survey of 196 participants without medical backgrounds using four fixed cases, not a real clinical workflow.

The authors note that all cohorts come from one geographic region, and differences between external cohorts suggest that equipment, acquisition protocols, populations and clinical practice may affect generalizability, so prospective validation across more institutions is warranted. Errors in upstream tooth localization, identification, lesion detection and segmentation can propagate through the staged architecture into later diagnostic descriptions and visual grounding. The patient-side evaluation used predefined cases and a web questionnaire, so effects on clinical efficiency, diagnostic decision-making, patient behaviour and clinical outcomes remain to be established in real-world studies. The system is positioned as an assistive tool, and outputs should be reviewed by qualified dental professionals before informing clinical decisions or patient care.