Research On CT Image Segmentation Of Liver And Liver Tumor Based On Deep Learning

Medical image segmentation is a hot research issue currently,it is a research field with multidisciplinary crossing,as well as an important application of computer graphics and image processing in biomedical engineering.Image semantic segmentation means that segmenting apart images and marking semantics on them,so as to enable images to accurately reflect information from each region.At present,on the issue of liver and liver tumor image segmentation,there are some problems such as unclear segmentation boundary,rough segmentation result,and excessive empirical parameters on traditional image segmentation method.Deep learning has the advantages of end-to-end,without designing feature rules and mining data potential features,which can well solve the above problems,therefore,the study on deep learning has profound research significance in image pixel-level segmentation.The study is divided into two major points,they are liver division and liver tumor division.Full supervised and weak supervised segmentation were used for liver segmentation,both of them have advantages and disadvantages: liver segmentation is more robust in a fully supervised learning mode;and the weakly supervised method can save the labeling cost and shorten the training time.In terms of liver tumor segmentation,a three-dimensional convolutional neural network with fusion characteristics is adopted focused on the characteristics of smaller liver tumors.This paper discusses the segmentation method of liver and liver tumor from the following three parts.(1)It builds a full convolutional semantic division network under full supervision and learning to introduce the attention mechanism to improve the network accuracy,accelerates the network training speed by changing the convolutional policy,at last,it analyzes network efficiency and accuracy from the perspective of math.The results show that the new network structure can maintain good accuracy while accelerating training.(2)For the problems of the high cost of time of liver segmentation,labeling involves professional knowledge,it introduces a method of the combination of deep learning and traditional segmentation under weak supervision,learning area of interest by migrating and attaining images through improving regional growth algorithm and image post-processing.Meanwhile,the performance of the fully supervised segmentation algorithm is compared mathematically.The experimental results show that the weak supervision is far higher than the full supervision learning method in terms of efficiency,and it can match the full supervision learning method in accuracy.(3)In terms of liver tumor segmentation,whose division difficulty is far greater than liver division due to irregularities in tumor shape,great differences in tumor intensity,and blurred boundaries between the tumor and the surrounding normal liver tissue.For this feature,this thesis adopts three-dimensional convolutional neural network with feature fusion,takes advantage of the three-dimensional convolution of interframe information in temporal dimensions,meanwhile,different characteristic layer of the simultaneous cascading network gives the network further attention to the small target areas in the images.The results show that three-dimensional convolutional neural network with feature fusion can effectively and continuously segment the liver tumor CT images.