Tumor Image Segmentation Method Based On Regional Fusion And Three-dimensional Convolutional Neural Network

Malignant tumor(cancer)has become one of the major public health problems that seriously threaten the health of the population in China.With the continuous development of medical Imaging technologies such as Magnetic Resonance Imaging(MRI)and Computed Tomography(CT),medical imaging can clearly observe the internal tissue structure and metabolic state of human body,which provide important reference information for modern doctors to make diagnosis.Tumor image segmentation refers to the accurate description of the tumor or tumor substructure,which plays an important role in the diagnosis and treatment of tumor,and can be used in tumor evaluation,preoperative program planning,postoperative effect evaluation and other aspects.With the increasing amount of data of clinical oncology,using computer to tumor image segmentation has become a hot research problem.At present,tumor image segmentation mainly faces the following problems :(1)there are similar topological relationships among tumor substructures of different patients,which have not been effectively mined by the existing methods;(2)tumor image usually has multiple modal image,the image of different modal organizations has different response to the human body,but this kind of multi-modal image has not been adequately used by the existing methods;(3)due to patient privacy and high cost of tumor image labeling,tumor image data with accurate labeling is small;To solve the above problems,this paper proposes a tumor image segmentation method based on regional fusion and three-dimensional convolutional neural network,with the following key contents: 1.Aiming at the problem of insufficient mining of topological relationships among brain tumor substructures in the current method,this paper proposes a method of brain tumor substructure segmentation based on hierarchical three-dimensional convolutional neural network.Firstly,the whole tumor was segmented by the convolutional neural network at the top layer,and the segmentation results of the region were obtained.Next,the result of multiplying the whole tumor segmentation results and the original input data is used as the input of middle layer convolution neural network,the middle layer of the convolution neural network output is the segmentation result of tumor core area.Finally,the result of multiplying original input data and the tumor core segmentation results is used as the input of bottom layer convolution neural network,then enhancement tumor region segmentation results are obtained.With this hierarchical segmentation strategy,the most peripheral tumor substructure can be used to limit the segmentation of internal tumor substructure.This method fully explored the inclusion relationship among different substructures of brain tumors and improved the accuracy of segmentation results.2.In order to solve the problem that the current method does not efficiently make use of multimodal brain tumor images,this paper proposes a method for substructure segmentation of brain tumors based on multimodal weight learning and three-dimensional convolutional neural network.This method takes advantage of the different characteristics of the imaging results of brain tumor images on human tissues in T1,T1 Gd,T2,and FLAIR modes,and weights the images of different modes to highlight or weaken the effect of images of certain mode on the segmentation results.As the parameter of the convolutional neural network,the weights of each mode can be dynamically learned according to the input data.The method makes use of the different characteristics of the brain tumor images with different modes to responding to the substructure of the brain tumor,and improves the accuracy of the segmentation results.3.To solve the problem that supervised learning methods are in great demand for tumor data with accurate labeling,this paper proposes an interactive three-dimensional lung tumor segmentation method based on multi-scale regional fusion.The method firstly requires users to make simple markers for lung tumors,on which the image patches of lung tumors is extracted and sampled at multiple scales.Then,the tumor image patches are segmented into different regions by using the method of super voxel method.Finally,the pre-segmentation regions are fused based on the principle of maximum similarity and adjacent.This method only requires users to make a simple mark on the tumor,and takes into account the threedimensional structure information of the tumor image.And the segmentation speed and accuracy of this method is fast and high.