Subcortical Brain Segmentation in MR image based on Residual Fully Convolutional Networks

Brain magnetic imaging(MRI) has become an effective tool for disease diagnosis, disease tracking and so on. Deep brain structures such as hippocampus plays an important role in the quantitative assessment of human brain. Therefore, it is crucial to segment these deep brain structures in a short time. In recent years, deep learning has achieved great process in image processing, especially the U-net, which is the Champion Model of 2015 ISBI Cell Tracking Challenge. In this paper, a novel structure was proposed which based on U-net as well as substitute the normal convolution block with the residual block to achieve a robust segmentation. In addition, SEblock is introduced to emphasize the correlation between channels which helps to focus on the features that are more important. The experimental results on the IBSR dataset show that the method proposed in this paper achieves average Dice value 0.83 on the deep brain structures segmentation, which verified the effectiveness of the method.