Populationbased generalizability of deep learning models in bitewing caries diagnostics

Status

Active, not recruiting

Phases

Unknown

Study type

Observational

Source

DRKS

Registry ID

DRKS00040737

Enrollment

1200

Registered

2026-06-23

Start date

2026-08-01

Completion date

Unknown

Last updated

2026-08-03

For informational purposes only — not medical advice. Sourced from public registries and may not reflect the latest updates. Terms

Conditions

Caries

Interventions

Group 1: Germany Model (G) Retrospective observational study arm evaluating a deep learning model for the detection of approximal carious lesions on bitewing radiographs. The model is trained exclusiv

Eligibility

Sex/Gender

All

Age

18 Years to No maximum

Inclusion criteria

Inclusion criteria: Patients in the permanent dentition. Availability of anonymized unilateral or bilateral bitewing radiographs acquired between 2013 and 2025 at either the Department of Operative Dentistry and Periodontology, University Medical Center Freiburg, Germany, or the Department of Conservative Dentistry and Endodontics, Saveetha Dental College and Hospitals, Chennai, India. Radiographs allowing complete assessment of at least one interproximal surface in the posterior dentition.

Exclusion criteria

Exclusion criteria: Incomplete image datasets. Radiographs of insufficient quality or otherwise unsuitable for diagnostic evaluation. Cases in which no complete assessment of at least one interproximal surface in the posterior dentition is possible.

Design outcomes

Primary

Measure	Time frame
Diagnostic performance of three deep learning models for the detection of approximal carious lesions on bitewing radiographs, assessed by positive predictive value (PPV), negative predictive value (NPV), sensitivity, and Cohen’s kappa coefficient. Outcome measures will be assessed once following completion of model training and evaluation on the retrospective study datasets. The deep learning models will be applied to anonymized bitewing radiographs from Freiburg, Germany, and Chennai, India. Model predictions will be compared with an expert-derived reference standard established by three calibrated dentists with more than ten years of clinical experience. Positive predictive value, negative predictive value, sensitivity, and Cohen’s kappa coefficient will be calculated to assess diagnostic performance and agreement with the reference standard.	—

Measure

Time frame

Diagnostic performance of three deep learning models for the detection of approximal carious lesions on bitewing radiographs, assessed by positive predictive value (PPV), negative predictive value (NPV), sensitivity, and Cohen’s kappa coefficient. Outcome measures will be assessed once following completion of model training and evaluation on the retrospective study datasets. The deep learning models will be applied to anonymized bitewing radiographs from Freiburg, Germany, and Chennai, India. Model predictions will be compared with an expert-derived reference standard established by three calibrated dentists with more than ten years of clinical experience. Positive predictive value, negative predictive value, sensitivity, and Cohen’s kappa coefficient will be calculated to assess diagnostic performance and agreement with the reference standard.

—

Secondary

Measure	Time frame
Diagnostic performance of the deep learning models as assessed by the F1-score, representing the harmonic mean of precision and sensitivity for the detection of approximal carious lesions on bitewing radiographs. The outcome measure will be assessed once following completion of model training and evaluation on the retrospective study datasets. The deep learning models will be applied to anonymized bitewing radiographs from Freiburg, Germany, and Chennai, India. Model predictions will be compared with an expert-derived reference standard established by three calibrated dentists with more than ten years of clinical experience. The F1-score will be calculated for each model to assess the balance between precision and sensitivity in lesion detection.	—

Measure

Time frame

Diagnostic performance of the deep learning models as assessed by the F1-score, representing the harmonic mean of precision and sensitivity for the detection of approximal carious lesions on bitewing radiographs. The outcome measure will be assessed once following completion of model training and evaluation on the retrospective study datasets. The deep learning models will be applied to anonymized bitewing radiographs from Freiburg, Germany, and Chennai, India. Model predictions will be compared with an expert-derived reference standard established by three calibrated dentists with more than ten years of clinical experience. The F1-score will be calculated for each model to assess the balance between precision and sensitivity in lesion detection.

—

Countries

Germany, India

Contacts

Public ContactDavid Auer

Klinik für Zahnerhaltungskunde und Parodontologie, Department für Zahn-, Mund- und Kieferheilkunde, Universitätsklinikum Freiburg

david.auer@uniklinik-freiburg.de+4976127048670

Outcome results

None listed

Source: DRKS (via WHO ICTRP) · Data processed: Aug 10, 2026