Research Of Brain Tumor Image Segmentation Method Based On Deep Learning

Brain tumor is a kind of brain disease threatening human health seriously.The patients suffer from tumor’s torture both physically and mentally.This disease causes high rate of mortality,too.Tumor segmentation based on MRI images bring great convenience to doctors in assessing the state of disease.In recent years,the rapid development of deep learning has greatly improved the accuracy of segmenting brain tumor,and deep learning gradually replaces traditional image processing methods and machine learning methods.Aiming at the problems of much time consuming and complex network scale in existing Dense Net methods,the DW-GN-Dense Net is proposed.Based on Dense Net,it replaces normal input convolution by dilated convolution with different dilation,and changes normal convolution with depthwise separable convolution,and uses group normalization instead of batch normalization,considering small number of samples in medical image processing task.Compared with original network,the modified network has better feature extraction ability,higher accuracy.In order to solve the problem that extracting training data window mismatch tumor in training strategy,a modified V-net called DW-SE-Vnet is proposed for preliminary tumor segmentation.The DW-SE-Vnet changes input convolution to depthwise separable convolution,and adds SEblock on original network structure.The DW-SEVnet has better performance in segmenting and lower time complexity.After obtaining preliminary tumor segmentation area by the DW-SE-Vnet,the DW-GN-Dense Net is trained with these data,promoting the DW-GN-Dense Net segmenting accuracy further.The cascaded network,yielding from combining the DW-SE-Vnet with the DWGN-Dense Net,has both advantages.Experiments show the cascaded network is more accurate than the DW-GN-Dense Net with silimar segmentation efficiency.Comparing with other methods,the proposed method can keep certain accuracy rate and less time taken by segmentation,and completes the segmentation task quickly.