A comparison of deep learning U‐Net architectures for semantic segmentation on panoramic X-ray images

Abstract

Digital image processing utilizes deep learning to tackle challenging issues such as image colourization, classification, segmentation, and detection. The medical image analysis field is developing day by day, and segmenting organs, diseases, or abnor malities is a challenging task to complete. Dental disease diagnosis is one of these fields where image segmentation can help gain significant improvements as dentists worldwide face various problems in diagnosing dental diseases with the naked eye. Compared to other medical images, dental radiographic images provide multiple challenges in terms of processing, making segmentation a more complex task. Deep neural network models are used more frequently for various image segmentation ap plications. U-Net is one such model. Multiple variations and advancements have been created for this network model to serve better performance, mainly on seman tic segmentation of medical images. However, comparative studies must determine how well these variants perform in segmenting dental x-ray images. This research uses six U-Net architecture (Vanilla U-net, Dense U-net, Attention U-net, SE U net, Residual U-net, R2 U-net) variants for segmenting dental radiographic X-rays that are extensively and effectively compared. Some U-Net architectural variations under consideration still need to be evaluated for segmenting dental radiographic X-rays. For all architectures, we used 2 and 3 convolutional layers. We used four types of matrices to compare the models: Accuracy, Dice coefficient, F1 score and IoU. Among the variants, Vanilla-unet with two convolutional layers provided the best Accuracy of 95.56% and IoU score of 88% on the validation set for much lesser time than other architectures. On the other hand, when we use three convolutional layers, dense-unet provides the best Accuracy of 95.94% and IoU score of 89.07% on the validation set. However, most of the examined architectures throughout the dataset showed minor changes when segmentation performance was measured using all four accuracy metrics. This study indicates that U-Net is enough for radio graphic X-ray segmentation. Choosing simpler models will save time and money during testing and model creation. Therefore, our suggested approach might aid in making automated dental disease diagnosis models.