| It is difficult to diagnose pancreatic cancer at an early stage,which is generally diagnosed as malignant.Automatic and accurate segmentation of the pancreas by computer for CT images is very important for the diagnosis and prognosis of pancreatic diseases.Due to the small volume,low contrast,and significant positional differences of the pancreas in CT images,pancreatic segmentation is difficult.It is a very important and challenging issue for segmenting pancreatic organs from CT images accurately and automatically.To address these issues,this thesis aims to achieve high-precision segmentation of the pancreas by constructing a deep learning model,and proposes a SAU-Net model based on attention mechanism and superpixel segmentation.(1)RAU-Net segmentation model is proposed firstly,where in the framework of U-Net,the convolutional blocks are placed by residual blocks,and a residual dual attention module is also added in the skip connection between the encoder and decoder.The module consists of RCM module and RSM module,which contains both of channel information and spatial information.At the same time,residual skip connections are added in the RCM module and RSM module,enable the model to learn high-resolution feature maps with spatial information.The average Dice coefficient of the RAU-Net model is up to 82.57%.(2)To solve the problem of unsatisfactory pancreatic edge segmentation,a superpixel segmentation module is plugged into the model and reaches the SAUNet model.The model is divided into two parts,the first step is to use the SLIC algorithm to perform superpixel segmentation on the image.The contrast of pancreatic boundaries in the image is enhanced based on the similarity of superpixel block features.The second step is to train the preprocessed image using the RAUNet model.From the model prediction results,the average Dice coefficient of the SAU-Net model is 85.02%,which is 2.45%higher than the RAU-Net model.The performance illustrates the priority of the SAU-Net. |
PDF Full Text Request