Analysis and prediction making on dental imagery using machine learning

• Applicant: Professor Dr. Falk Schwendicke, since 8/2022
• Subject Area: Dentistry, Oral Surgery
• Term: from 2020 to 2024
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 445925495

Dental diseases like caries, periodontitis or apical inflammations (apical lesions) are among the most prevalent conditions in the world, generating significant burden to individuals and healthcare systems. Dental imagery data is an indispensable source for detecting and assessing dental diseases. In particular in dentistry imaging is abundant (>50% of all radiographs taken in Germany originate from dentistry) and inherently multi-modal (intra- and extra-oral radiography, CBCT, transillumination, photography, among others, often from the same patient). Combined with clinical patient records, such a diverse image repository allows for cross-sectional and longitudinal analysis, risk analysis and prediction making. However, having a human expert integrating these multiple data sources is close to impossible. We aim to combine different image modalities and (semi-) structed text-based patient related health data, something we coin as "cross-talk", to evaluate the efficacy of different machine learning (ML) techniques. ML was applied successfully in healthcare. In dentistry, however, ML has only been sparsely applied. In previous studies we applied convolutional neural networks (CNNs), a specialized kind of ML, which is particularly useful to extract hierarchal features from image data, to classify radiographic images of teeth for periodontal bone loss and apical lesions, with the diagnostic performance of CNNs being similar to average dentists. We further applied CNNs to discriminate carious from non-carious teeth based on near infrared-light transillumination (NILT) images, again with reasonably well performance despite being trained on a small dataset. We further applied CNNs for tooth detection and tooth classification on panoramic radiographs as well as on bitewing radiographs, as well as pixel-wise segmentation of teeth, restorations and tooth structures (enamel, dentin and pulpa). While these studies focused on eligibility and were performed on small data set sizes, they highlighted the potential of ML/CNNs to serve the needs of clinicians and patients. As a consequence, we at the dental hospital at Charité – Universitätsmedizin Berlin built a data base of clinical records (> 20,000) and annotated image materials (> 10,000), a prerequisite to apply advanced ML. In the planned studies, we will – within a hypothesis-driven analytic framework – (1) assess if the integration of different image modalities allows to overcome the shortages of one particular imaging technique, hence outperforming human experts; (2) evaluate strategies to deal with class imbalances, an inherent challenge of data sets originating in the life sciences; and (3) assess the potential of longitudinal prediction making by combining clinical records and imagery data. The planned research will contribute to a better understanding of the application of CNNs in dentistry. The research will help to leverage AI-based technologies in dentistry to benefit clinicians and patients.

• DFG Programme: Research Grants
• Co-Investigators: Professor Dr. Christof Dörfer; Professor Dr. Sebastian Paris
• Ehemaliger Antragsteller: Dr. Joachim Krois, until 8/2022