Intelligent biofilm detection with ensemble of deep learning networks.

Abstract

Background: Dental biofilm is traditionally identified visually, which can be challenging and time-consuming due to its color similarity with the tooth. The aim of this study was to evaluate the performance of U-Net neural networks for the automatic detection of dental biofilm without disclosing agents on intraoral photographs of deciduous and permanent teeth using an ensemble strategy.

Material and methods: This retrospective exploratory study was conducted on two datasets of intraoral images obtained from deciduous and permanent dentitions. The first dataset was used to validate dental biofilm annotations by an expert with disclosing agents. The second dataset, without disclosing agents, was employed to train and evaluate the U-Net neural network in the identification of dental biofilms using an ensemble strategy.

Results: The performance of the ensemble method was assessed using a cross-validation procedure, with six groups dedicated to training, one group for validation, and one group exclusively taken as a test set for the final evaluation of the ensemble. The performance of the neural network was evaluated using accuracy, F1 score, sensitivity, and specificity. The U-Net neural network achieved an accuracy of 93.1%, sensitivity of 65.1%, specificity of 95.9%, and an F1 score of 63.0%.

Conclusions: The U-Net neural network using the ensemble strategy was able to automatically identify visually detectable dental biofilms on intraoral photographs. The application of this new knowledge will soon be available in clinical practice.