Segmentação automática da placa dentária com base em aprendizado profundo

Abstract

Dental plaque biofilm is the main etiologic factor for dental caries and periodontal diseases. However, its visualization is difficult, and the use of disclosing agents is a laborious and unpleasant process. Therefore, plaque identification through an automatic process is important. The present research aimed to apply a Convolution Neural Network (CNN) model to segment dental plaque in intraoral digital photographs without the use of disclosing agents. The dataset used to evaluate the proposed system included 480 intraoral photos including lateral and frontal views of permanent and deciduous dentition, contemplating the presence and absence of orthodontic appliances. The photographs were divided into three subsets: 360 images were used for training; 60 photos were used for validation; and 60 photos were used for testing. All images have been labeled by a specialist dentist with over 30 years of experience. The U-Net architecture was used for image segmentation. Metrics of accuracy, sensitivity, specificity and F1 score were used to evaluate the performance of the model in each dental unit. The trained model obtained 91.8% accuracy, 67.2% sensitivity, 94.4% specificity and 60.6% F1 score. These metrics were chosen for their easy interpretability (accuracy), their use in health areas (sensitivity and specificity) and for weighting unbalanced classes (F1 score). A higher plaque fraction was observed in the lateral view images, as well as in the images with orthodontic appliances. These images also exhibited higher F1 scores (61.7% and 61.5%, respectively) and specificity (94.5% and 95.6%, respectively). In conclusion, a deep learning method for segmenting dental biofilm in permanent and deciduous dentitions is feasible and could be a visual aid tool to improve oral hygiene and patient control of dental plaque.