Modified U-Net For Brain MR Image Segmentation

Cardiovascular and cerebrovascular diseases are serious frequently-occurring diseases.Early diagnosis and treatment are of great significance to prevent the diseases.It’s a common diagnosis measure to use image segmentation technology to analyze medical images,compared with other imaging techniques,images from MRI are easier to be used for diagnosis of brain disease for their high contrast among different soft tissues and high spatial resolution,and MRI is non-invasive.However,due to the problems of imaging equipment and the particularity of brain tissues,there are many problems in brain MR images,such as noise,weak edges and intensity inhomogeneity.Traditional image segmentation algorithms are mostly based on gray value,which causes it difficult for them to obtain accurate segmentation results.By comparison,models based on convolutional neural networks(CNN)like fully convolutional networks(FCN)and U-Net perform better,but the segmentation results of details are not good enough and the boundaries are smooth because of so many convolution and pooling operations.Based on U-Net,a multi-scale U-Net(MSFCN)is proposed in this paper.To improve the performance of dealing with details and edges,small scaled filters are used in MSFCN,meanwhile,big scaled filters are used to expand the receptive field to utilize more neighborhood information.The performance of the model is improved by combining the feature maps of different scaled convolution filters.Based on cross-entropy(CE),a weighted cross-entropy loss function is constructed by introducing weight coefficients.Total loss is calculated by summing single CE pixel by pixel,but there is a large area of background in medical images and the proportion of different tissues are unbalanced,which cause the segmentation results of background influence loss a lot.The weighted cross-entropy loss function is used to balance the effect of different classification results on target loss by controlling the pixels’ weights,so as to improve the segmentation accuracy of the model.