Determination of developmental tongue anomalies using convolutional neural networks

The aim of this study is to identify tongue anomalies that encountered during routine dental examination using convolutional neural networks. 1038 tongue photographs were included in the study and classified into 6 classes as: 123 healthy tongue, 259 coated tongue, 289 fissured tongue, 96 hairy tongue, 174 geographical tongue and 97 median rhomboid glossitis. A binary classification process was performed between healthy tongue and other tongue anomalies. In addition, four different triple classification problems were evaluated as: healthy-coated-fissured, healthy-hairy-fissured, healthy-hairy-coated and coated-hairy-fissured tongue. The ResNet18 model is preferred for binary classes; ResNet50 model was preferred for triple classes. The performance of the models was evaluated with accuracy, recall, precision and F1-Score metrics. The ResNet18 model generally demonstrated high performance for each binary class in the binary classification problem. Especially for the healthy-hairy language group, 100% success rate was achieved in all metrics. When other groups were analyzed, high metric values were observed. This shows that the ResNet18 model is effective in binary tongue classification problems on the data used within the scope of the study. The ResNet50 model was preferred for more complicated triple classification tasks. High accuracy (0.96) and F1-Score (0.95) values were obtained for the healthy-coated-hairy tongue group. It was observed that the model achieved a relatively high level of performance (0.91 accuracy and 0.91 F1-Score) for the fissured-coated-hairy tongue group. The model could not provide the same performance for the healthy-fissured-coated and healthy-fissured-hairy tongue groups. This indicates that these anomalies are more difficult to distinguish and the model struggles with complex tasks where differantiation is challenging. With convolutional neural networks, automatic diagnosis of language anomalies, which are often asymptomatic and overlooked, seems possible. Algorithms developed through more comprehensive studies will help clinicians with a more efficient diagnostic process.