A Deep Learning Algorithm Based on U-Net for Automatic Tongue Region Segmentation from Images Acquired through a Computerized Tongue Image Acquisition System

Research on improving the accuracy of tongue region segmentation in the deep learning process of Korean medicine's tongue diagnosis is actively ongoing. This study aims to propose a segmentation model based on the U-Net using Convolutional Neural Networks (cNN). Binary mask images representing the tongue region were created, and the dataset was randomly divided into an 8:2 ratio for the train and test set. To prevent overfitting, the changes in validation loss were monitored at each epoch. Finally, Dice similarity coefficient (DSC) and Jaccard index (JI) were computed to measure the accuracy of the segmentation model in identifying the target regions. The modified U-Net network showed a significant level of accuracy in tongue region segmentation and minimal overfitting despite a small number of training data sets. The accuracy in the train set reached 0.997 and in the test set the accuracy was 0.993. The Dice similarity coefficient (DSC) score on the test dataset was 0.981 ± 0.017 and the Jaccard index (JI) score was 0.964 ± 0.031. The proposed model based on U-Net for tongue region extraction is anticipated to be effective for practical applications such as computerized tongue analysis systems.