Pancreas Segmentation In Abdominal CT Scans Based On Deep Convolutional Neural Network

A large number of multi-view and multi-position abdominal computed tomography(CT)images can be obtained by CT technology.Accurate pancreas organ segmentation based on abdominal CT scans is the prerequisite and key step of computer-aided system for disease diagnosis,medical image analysis and surgical planning of tumor ablation.However,pancreas segmentation is challenging because of the characteristics of blurred boundaries and high gray-scale similarities between adjacent organs in abdominal CT images,as well as the high variability in the position,shape and size of pancreas organ among different patients.The excellent performance of deep convolutional neural network in the field of image segmentation demonstrates its super learning ability and provides a new research idea for pancreas organ segmentation in abdominal CT scans.Therefore,this thesis focuses on pancreas organ segmentation based on deep convolutional neural networks:(1)Pancreas organ has very different scales and shapes in CT slices,and require a receptive field of appropriate size to achieve accurate segmentation.Therefore,this thesis uses k U-Net network,through the cooperation of two U-Net subnetworks on different scales of CT slices,to complete the multi-scale context information extraction within CT slices.On the other hand,in medical CT images,there is spatial sequence context information between adjacent slices.Thus,on the basis of using k U-Net network to extract the intra-slice context information,this thesis further designs a bi-directional convolution gated recurrent unit(BDC-GRU)network to extract the inter-slice context information.The experimental results show that the algorithm combining k U-Net network and BDC-GRU network can effectively improve the accuracy of pancreas organ segmentation,and the dice similarity coefficient(DSC)index reaches 87.63%.(2)The proportion of pancreas organ in the whole CT slice is small,resulting in the existing fully convolution network difficult to learn effective information,so that the accuracy of pancreas organ segmentation is not very ideal.To solve this problem,the hierarchical supervised pyramid network is proposed in this thesis.Through the encoder and decoder structure with jump connection,the pyramid network can combine the shallow features with high resolution and the deep features with strong semantic information to form pyramid semantic features with rich information.Finally,through the introduction of hierarchical supervision,the pyramid network can fully learn effective information in the process of training.Sufficient experiments show that this algorithm can achieve accurate segmentation of pancreas organ,and the DSC index is increased to 88.28%.(3)The boundary of pancreas organ is tortuous and changeable.In order to further improve the segmentation accuracy of pancreas organ,it is necessary to ensure that the boundary areas of pancreas organ which are difficult to be segmented can be segmented accurately.Therefore,on the basis of pyramid network,a Fine-grained Refine Net is designed to realize the fusion of pyramid semantic features.Finally,a pancreas organ segmentation algorithm based on pyramid network and Refine Net is proposed to modify the results of pancreas organ segmentation.The experimental results show the effectiveness of the refined network,and the segmentation algorithm improves the pancreas organ segmentation accuracy by 0.48% based on the pyramid network,and the DSC index reaches 88.76%.