Automatic Brain Tumor Segmentation Method Based On Deep Learning

Glioma is a common malignant brain tumor,which has the characteristics of irregular shape,fuzzy boundary and different degrees of invasion.Because of the specificity and complexity of human brain,it is diffcult to segment tumors accurately from multimodal brain MRI images.Compared with normal brain tissues and background area,tumor areas account for a small proportion,which leads to the serious imbalance of foreground and background pixels in the dataset.When using deep learning framework to learn target features,it is easily affected by background features,so it is impossible to train models effectively.In order to solve the problem of imbalanced categories in brain tumor image segmentation,two methods are proposed in this paper:1.Improved two stage automatic brain tumor segmentation algorithm based on Mask RCNN.The algorithm is divided into two steps:(1)extracting features and generating candidate frames from preprocessed brain images by using regional convolution neural network.(2)The improved full convolution neural network with attention mechanism module is used to infer the generated candidate box,predict its category,bounding box regression position and segmentation maps.The attention mechanism makes the model focus on the tumor pixels during training,and effectively suppresses noise and background information.The Bra TS19 dataset provided by brain tumor challenge is used to test the performance of the improved algorithm in brain tumor image segmentation.Compared with the Mask RCNN model,the improved Mask RCNN model improves the Dice coefficient and sensitivity by 1%～2%.The results prove that the introduction of attention mechanism module can improve the accuracy of tumor segmentation with only a little more computation.2.An algorithm for brain tumor segmentation based on nested U-Net network model with focus loss function.The convolution layer of each model of U-Net is replaced with a small U-shaped structure in our proposed algorithm can increase the depth of the model and capture more information in the stages when the amount of computation is increased.In order to obtain a fine segmented prediction map side auxiliary output structure is added to fuse the prediction map output in the decoding stage.The focus loss function is used to make the model pay more attention to the sample points which are few and difficult to classify in brain images during training,so as to solve the category imbalance problem in brain MRI images.In the testing stage,Dice coefficients of all tumors,tumor core and enhanced tumors were 0.854,0.873 and 0.791,respectively.Comparing the prediction results of two algorithms,it shows that for brain images with simple semantic information,the single stage nested U-Net can extract enough brain tumor features efficiently,and achieve higher segmentation accuracy.