Abstract:
This independent study aims to develop a model for segmenting proximal dental caries using a fully convolutional neural network in bitewing radiographs. The segmentation models were created with the explicit goal of helping dentists in segmenting dental caries in radiographs for a second opinion. To determine the most appropriate model architecture, we compared the performance of three fundamental segmentation models: U-Net, FPN (Feature Pyramid Network), DeepLabV3+, and XsembleNet, which is a combination of the three preceding models. The system is evaluated in two ways. The first is to assess segmentation quality using the dice coefficient ; empirical experiments indicate that XsembleNethas the highest dice coefficient, followed by FPN. The second evaluation is to rate models segmentation of 12 testing bitewing radiographs. While all four models are comparable in terms of accuracy and specificity, XsembleNet and FPN jointly achieve the highest classification metrics score. As a result, it can be concluded that a fully convolutional neural network could be used to detect dental proximal caries radiographs via computer assisted diagnosis.