Pathology detection on third molars using deep learning

Thumbnail Image
Issue Date
Journal Title
Journal ISSN
Volume Title
Objective The aim of this study is to automatically locate the third molars(M3) on a presurgical panoramic radiograph (PR) and assess the presence of pathologies on or around each M3. Materials and methods 2457 PRs were manually annotated and labelled. A segmentation model for each M3 was trained using a MobilnetV2 backbone. These segmentations were used to create a Region of Interest (ROI) crops for three different pathologies: cysts, caries- and periapical lesions. A classification network was trained for each of these pathologies. The outputs of all models were subsequently combined into a summarized view. Results The segmentation models reached F1-scores ranging from 0.89 to 0.91. Caries lesion classification reached an accuracy of 0.86. An accuracy of 0.95 was achieved on Periapical lesion classification. For the classification of cysts an accuracy of 0.63 was achieved. Conclusion Automated assessment of PRs prior to third molar removal can assist the assessing clinician by reducing manual labour and interrater variability and by giving them a robust indication of pathologies present on or around the third molars. Keywords: deep learning, artificial intelligence, panoramic radiograph, third molar, caries, periapical lesion, cyst
Faculteit der Sociale Wetenschappen