A Comparative Study of U-Net-Related Networks Under Dental X-Ray

In the field of dental pathological diagnosis, designing deep learning algorithms to extract and automatically segment regions of interest in dental X-ray images simplifies the difficulty of interpretation for dentists and will become an important tool for auxiliary diagnosis in dentistry. UNet has been widely favoured by scholars since its design due to its encoding-decoding design and coarse- and fine-grained feature fusion, however, very few people have done a detailed analytical study of U-Net and its variants, which makes the medical field and image engineering lack of selection of Unet and its typical improved algorithms to be used, the aim of this study is to make a detailed comparative study of UNet, and related structural variants of UNet to do a detailed comparative study. Specifically, the logical structure UNetCT with convolutional downsampling and transposed convolutional upsampling is used as the basic UNet to explore the variant structures of UNet++ with nested and dense skip pathways, ResUNet with residual connections, and DenseUNet with densely concatenated blocks. It is found that the simple structure UNet and the residual structure ResUNet in dental X-Ray achieved better results, while the more complex structure UNet++ achieved worse results with a large number of parameters, and the DenseUNet achieved better results and the smallest number of parameters on the training set, but the worst results on the validation set. Our research solves the difficulty of applying UNet and UNet variants in industry and biomedicine to bring practical references, and the successful segmentation of teeth also brings convenience and efficiency to computerized automatic segmentation technology to assist medical diagnosis.