Research On MRI Brain Tumor Segmentation Algorithm Based On Super-pixels And Fully Convolutional Neural Network

Brain tumor is an abnormal tissue hyperplasia that will damage the central nervous system and even threaten the patient’s life.Magnetic resonance imaging(MRI)has high soft tissue resolution,so MRI technology is commonly used in clinical applications.Brain tumors are usually segmented by radiologists,but this method is laborious and time-consuming and prone to errors,and the results are difficult to reproduce.The computer-aided tumor segmentation method realizes the automatic segmentation of tumors,which not only saves time but also achieves repeatable results.Therefore,computer-aided tumor segmentation plays a vital role in the field of modern medicine.This topic has designed a feasible brain tumor segmentation method:first,the four-modal brain tumor image is super-pixel segmented for the first time to obtain the region of interest(ROI);the second step is to use ROI as a fully convolutional neural network(FCN)input is processed to get the results of each mode;the third step is to multi-label fusion of the four modal segmentation results to obtain the final result.This article uses data from the Brain Tumor Segmentation International Challenge(BRATS)2017.In this paper,the MRI brain tumor is divided into three internal tissues,namely edema area(ED),active tumor area(ET)and necrosis area(NCR).Using similarity coefficient(DICE),specificity(TPR)and sensitivity(PPV)evaluation coefficients to quantify the entire tumor area(WT)and tumor core area(TC),the evaluation results show that its DICE is 0.98 and 0.94,respectively,PPV were 0.99 and 0.94,and TPR were 0.99 and 0.99,respectively.The experimental results show that the proposed MRI brain tumor segmentation algorithm based on the combination of superpixels and FCN can get better results.Using the ROI obtained by the superpixel method as the input of FCN can make the segmentation results more accurate.At the same time,a comparative experiment was conducted.When the entire image is directly used as the input of the FCN without using ROI.the accuracy of the output result will be worse.For example,some small tumor regions on the edge are not accurately segmented.Through comparative analysis,it can be concluded that the MRI brain tumor segmentation method based on superpixels and fully convolutional neural networks shows relatively high performance in the segmentation of brain tumor subregions.The algorithm proposed in this paper provides a new method for automatic segmentation of multimodal MRI brain tumor images,and has certain clinical reference value in the diagnosis and treatment of brain tumors.