Detecting white spot lesions on dental photography using deep learning: A pilot study. (April 2021)

• Record Type: Journal Article
• Authors: Askar, Haitham; Krois, Joachim; Rohrer, Csaba; Mertens, Sarah; Elhennawy, Karim; Ottolenghi, Livia; Mazur, Marta; Paris, Sebastian; Schwendicke, Falk
• Is Part Of: Journal of dentistry. Volume 107(2021)
• Publication Date: 2021-04
• Subjects: Artificial intelligence -- Caries -- Digital imaging/radiology -- Mathematical modeling -- Photography -- White spots; Dentistry -- Periodicals; Dentistry -- Periodicals; Dentisterie -- Périodiques; Electronic journals; 617.6005
• DOI: 10.1016/j.jdent.2021.103615
• Languages: English
• ISSNs: 0300-5712

Abstract

Abstract: Objectives: We aimed to apply deep learning to detect white spot lesions in dental photographs. Methods: Using 434 photographic images of 51 patients, a dataset of 2781 cropped tooth segments was generated. Pixelwise annotations of sound enamel as well as fluorotic, carious or other types of hypomineralized lesions were generated by experts and assessed by an independent second reviewer. The union of the reviewed annotations were used to segment the hard tissues (region-of-interest, ROI) of each image. SqueezeNet was employed for modelling. We trained models to detect (1) any white spot lesions, (2) fluorotic lesions and (3) other-than-fluorotic lesions. Modeling was performed on both the cropped and the ROI images and using ten-times repeated five-fold cross-validation. Feature visualization was applied to visualize salient areas. Results: Lesion prevalence was 37 %; the majority of lesions (24 %) were fluorotic. None of the metrics differed significantly between the models trained on cropped and ROI imagery (p > 0.05/ t -test). Mean accuracies ranged between 0.81−0.84, without significant differences between models trained to detect any, fluorotic or other-than-fluorotic lesions (p > 0.05). Specificities were 0.85−0.86; sensitivities were lower (0.58−0.66). Models to detect any lesions showed positive/negative predictive values (PPV/NPV) between 0.77−0.80, those to detect fluorotic lesions 0.67 (PPV) to 0.86 (NPV), and those to detect other-than-fluorotic lesions 0.46 (PPV) to 0.93 (NPV). Light reflections were the main reason for false positive detections. Conclusions: Deep learning showed satisfying accuracy to detect white spot lesions, particularly fluorosis. Some models showed limited stability given the small sample available. Clinical significance: Deep learning is suitable for automated classification of retro- or prospectively collected imagery and may assist practitioners in discriminating white spot lesions. Future studies should expand the scope into more granular multi-class detections on a larger and more generalizable dataset.