Periodontitis is a major cause of tooth loss, yet sub-gingival breakdown and radiographic bone loss are still assessed manually with inter-clinician variability. This project will develop PerioTwin, a multimodal AI system that: (i) detects and stages periodontitis at tooth level from panoramic radiographs (OPG) supported by targeted intraoral photos; (ii) quantifies bone support via clinically meaningful landmarks (cemento-enamel junction, alveolar crest, root apex) to compute bone-loss ratios with calibrated uncertainty; and (iii) provides tooth-level risk stratification to improve screening and follow-up planning beyond static classification.
The project targets a clinically meaningful dataset (>300–500 OPGs) with clinician agreement ? > 0.75 on tooth-level staging (adjudicated subset). Target performance on a held-out test set is AUC > 0.90 (or macro-F1 > 0.80) for tooth-level detection/staging, MAE < 0.10 for bone-loss ratio estimation, calibration ECE < 0.05, and robustness with <10% relative performance drop across image-quality/device strata.
LSMU will provide anonymized (GDPR-compliant de-identified) patient imaging data (OPGs and targeted intraoral photos) and available periodontal indicators, and will lead the annotation protocol, QC, and clinical validation. KTU will develop tooth instance + landmark models, attention-based multimodal fusion, calibration/uncertainty, and reproducible evaluation. Deliverables include an annotated dataset + guideline, validated models with clinician-interpretable overlays, and a software prototype with reproducible evaluation and standardized reporting.
Project funding:
KTU fund for internal investment
Project results:
Target indicators (quantitative success criteria):
• Dataset + annotation: curate and de-identify a Lithuanian clinical dataset (target ?300–500 OPGs) plus targeted intraoral photos where available; achieve inter-rater agreement ? ? 0.75 for tooth-level staging on an adjudicated subset.
• Tooth-level detection/staging: on a held-out test set, achieve AUC ? 0.90 (or macro-F1 ? 0.80) for tooth-level disease detection/staging (depending on label scheme).
• Bone-support quantification: achieve MAE ? 0.10 in bone-loss ratio estimation (or equivalent error in the chosen metric) vs clinician reference; provide uncertainty estimates with ECE ? 0.05 calibration error.
• Robustness: performance drop across acquisition variability (device/quality strata) kept within ?10% relative vs in-distribution results, documented with failure-mode analysis.
• Additional reported metrics: report accuracy, sensitivity, and specificity at a predefined operating threshold for classification, and RMSE alongside MAE for bone-loss quantification.
Period of project implementation: 2026-04-01 - 2026-12-31
Project partners: Lithuanian University of Health Sciences
