CT Medical Image Segmentation Based On Cascaded Convolutional Neural Network

Medical image segmentation is an important research direction of computer vision and medical image analysis and has important application prospects in clinical diagnosis and treatment.Among them,liver tumor segmentation and pancreas segmentation are the difficult points in all medical image segmentation.The reasons include that liver tumor and the pancreas are both small targets in abdominal CT.The average volume of liver tumors occupies less than ten thousand of the volume of the entire abdominal CT.In addition,the variability of the internal structure of liver tumors results in unstable texture characteristics,which intensifies the difficulty of segmentation.As a normal organ,the pancreas is not as difficult to segment as liver tumors,but as one of the smallest organs in the abdomen,its segmentation also faces difficulties caused by small targets and unclear texture edges.At present,for the segmentation of small targets in medical images,the most commonly used solution is the cascade method,which is also applicable to the segmentation of liver tumors and the pancreas.Since the scientific problems and difficulties faced by liver tumors and pancreas in segmentation are similar,so do the solutions,this thesis combines the two types of segmentation tasks for discussion.Therefore,in this thesis,for the segmentation of liver tumor and pancreas in abdominal CT,a series of improved methods are proposed to achieve better segmentation accuracy for small targets.First of all,for the segmentation task of liver tumors,this thesis proposes a new Cascaded Deeply Supervised(CDS)convolutional neural network model that combines a deep supervision mechanism and a cascade segmentation method.This model uses the output of multiple sub-networks of the deep supervision model to realize the cascaded segmentation process in a single network.The CDS model can adaptively allocate pixels in the liver to sub-networks of different depths for classification.Among them,the shallow sub-network is responsible for classifying most of the easiest pixels,including tumor pixels and normal liver pixels.The deeper sub-network is responsible for further classification of pixels that the shallow sub-network cannot reliably classify.Using such a cascaded pixel classification mechanism,the output from the shallow layer to the deep layer is gradually completed,and finally,the entire network is covered,and the classification of all pixels is completed.The advantage of the CDS model is that it can effectively suppress the under-segmentation problem caused by the small tumor volume in the liver,thereby improving the accuracy of liver tumor segmentation.Second,in the task of pancreas segmentation,this thesis first proposes a new“Segment And Pickup(SAP)” framework to address the deficiencies of the commonly used “Locate And Segment(LAS)” framework.The SAP framework contains two independent pancreas segmentation networks.The first is the precise segmentation network of the pancreas,which is only responsible for learning the pixel information sampled near the pancreas.This network can accurately segment the pancreas,but it is easy to misclassify pixels far away from the pancreas.The second is the seed network,which is responsible for learning to classify all pixels in a complete CT.Although it does not misclassify pixels far from the pancreas,it is easy to under-segment the pancreas.The SAP framework uses the region growing method,taking the segmentation result of the seed network as the growth seed,and extracting the precise pancreatic segmentation result from the segmentation result of the segmentation network.The advantage of SAP is that it combines the advantages of two independent networks,and the final segmentation result is more accurate than any single network,and more importantly,better than the segmentation result of the commonly used LAS framework.Third,although the SAP framework can obtain accurate pancreatic segmentation results,it needs to train two segmentation networks independently,so the training time is long,which is not conducive to the rapid iterative optimization of the model.This paper further proposes a self-pickup deeply supervised(SPDS)model.The model can realize the functions of the segmentation network and seed network at the same time in one network and can achieve the pancreas segmentation accuracy of double networks in SAP.At the same time,this paper proposes a dual training method(DTM)for SPDS network training,which can make the SPDS network have the advantages of "precise segmentation close to the pancreas" and "error-free segmentation far away from the pancreas" in one training process,which significantly improves the efficiency of network training.Finally,this thesis proposes the expandable convolutional neural networks(ECNN)model to solve the difficult problem of 3D complex network training.The ECNN model adopts the expandable convolution kernel(ECK),and the training data of multiple scales is used in the training process to sub-join the training network model.The ECNN network with ECK only uses standard training in the cascade training process.Set,do not need to use an additional database to pre-train the model.ECK can speed up the training speed of the network.In addition,the pancreas segmentation performance of the ECNN network model is better without using the pre-training model.The liver tumor and pancreas segmentation methods proposed in this thesis have been tested on Li TS and NIH datasets respectively.Experimental results proved that the new method proposed in this paper can improve the accuracy and efficiency of liver tumor and pancreas segmentation.It is quite competitive when compared with the best-performing methods in the past.However,it will be used in the future application of three-dimensional networks and large-scale There is still room for improvement in network optimization.