Research On Parallel CNN Multi-scale Feature For Brain Tumor MRI Segmentation

Brain tumor image segmentation is an important branch of medical image processing.The aim is to obtain accurate information such as location,size and shape of tumor by intelligent processing of intracranial image data,so as to provide reliable basis for diagnosis for oncologists.Glioma is the most common central nervous system disease in the diagnosis of brain tumors,accounting for about 80% of primary tumors,and is commonly seen in children and young adults.The high grade glioma is characterized by variable morphology and infiltrating growth.For patients with space-occupying tumor compression,craniotomy may cause nerve damage to the patient,so experts usually make a diagnosis based on brain imaging data.In the intelligent non-invasive checking algorithm,the prediction results of the model built solely on the characteristics of strength threshold and texture feature are less accurate than the real tags,and it is difficult to be widely used in practice.Therefore,it is necessary to develop a reliable automatic segmentation algorithm.In this paper,two deep learning models are designed for the task of lesion segmentation of brain tumor images,namely the encoder-decoder structure based on u-net model and a two-channel 3d parallel convolution network.The structure of encoder and decoder has been proved to have good performance in multiple fields of medical image segmentation.Therefore,based on the u-net model,the spatial structure of 3d data of MRI image is constructed in this paper,which is tested in BraTS data set,and the Dice coefficient is calculated and compared with other full convolution algorithms.The good performance of u-net in image semantic segmentation is attributed to two points.First,the contraction path of u-net can obtain the context information in the image,while the expansion path can obtain the local features.Second,for some tumor data,the model after u-net training can locate the lesion location.In order to adapt to the limitation of computer video memory,this paper randomly extracts multi-scale image blocks as the input of the network in the sampling stage to complete data enhancement and solve the problem of less trainable labeling data.The second algorithm consists of an empty convolutional network and a dense fully connected network.The parallel convolutional network first extracts multi-scale image blocks for training,captures a large range of spatial information,and USES the identity mapping characteristics of dense connections to superimpose shallow features on the end of the network,and divides the edema area,enhancement area,non-enhancement area and encapsulated area in the MRI multi-mode image.In BraTS 2018 experiments to test the model validation data segmentation result of whole tumor,core,and enhance the average Dice is about 0.90,0.73 and 0.71,respectively,in the experiment over the emptiness of single channel network and accuracy of the segmentation results of dense to connect to the Internet,and published in the conference doing quite good algorithm,and has high automation integration.