Automatic Segmentation Research Of Liver Tumor In Abdominal Ct Based On Deep Learning

Computed tomography(CT)is the most commonly used image modality in clinical liver tumor evaluation.CT examination is widely used in clinical diagnosis and review due to its fast and non-invasive detection.It is time-consuming and laborious for doctors to analyze CT images manually.Therefore,in the field of computer-aided diagnosis,the automatic segmentation of CT images is a research hotspot.In CT image,liver tumor suffers from the giant difference in pixel value and the blurred edges,which makes it more difficult to get segmentation result automatically.With the rapid development of deep learning methods,the segmentation network based on the full convolutional network structure can provide doctors with pixel-level classification information and accurately locate organs and lesion areas.Compared with the two-dimensional convolution,the three-dimensional convolution structure can better capture the three-dimensional context characteristics of the image,and the segmentation network based on the three-dimensional full convolution network can more accurately segment the organ and the diseased area.The three-dimensional segmentation network requires high computing resources,so it is difficult to design a very deep network structure.In addition,since the slice spacing of abdominal CT scans is 5 mm,CT images lack three-dimensional continuity information.The resolution anisotropy of CT image will limit its segmentation effect in the three-dimensional convolution structure.In view of the above challenges,this thesis improves the three-dimensional cascaded segmentation network,and reconstructs the inter-layer information of CT images through the superresolution network to further improve the segmentation accuracy.The main contributions include:(1)Aiming at improving the segmentation network model,the advantages of different network structure improvement methods are discussed.Based on the idea of coarse-fine segmentation in the cascade network framework,an improved three-dimensional cascade U-Net segmentation network architecture is designed.(2)Focusing on the characteristics of the anisotropic resolution of CT images,the advantages and disadvantages of interpolation methods and super-resolution networks for image reconstruction are discussed.A three-dimensional super-resolution network based on the averaging module is designed to get better reconstruction of inter-layer information and finally improve the anisotropic resolution of CT images.(3)By combining the super-resolution network with the segmentation network,the superresolution network is introduced to reconstruct the CT image data and labels at the same time.After the reconstruction,improved three-dimensional cascade U-Net segmentation network architecture is used.By optimizing the anisotropic resolution of CT images and the segmentation network structure at the same time,the final segmentation accuracy is further improved.The comparative analysis of experimental results proves that the cascaded liver tumor segmentation network and the three-dimensional super-resolution generation adversarial network based on the average block designed by this paper can improve the accuracy of tumor segmentation.Firstly,the improvements of the cascade network allow the network structure designed in this paper to achieve better tumor segmentation results than existing methods on two different datasets.By comparing the interpolation method and different super-resolution networks,the three-dimensional super-resolution generation adversarial network based on the average block can restore the inter-layer information better and further improve the tumor segmentation accuracy.