Automatic Brain Tumor Segmentation Using Cascaded FCN with DenseCRF and K-means

Abstract

Accurate and reliable brain tumor segmentation is a critical component in cancer diagnosis and treatment planning. Based on the successful convolution neural network (CNN) technique for complex and high dimensional input problems, the fully convolution neural network (FCN) and fully connected conditional random field (DenseCRF) are applied to the segmentation of brain tumors. The K-means and FCN are combined to improve the segmentation accuracy. DenseCRF is used to optimize the FCN segmentation results, and then the model fusion is used to further supplement the segmentation results to improve the segmentation accuracy. The experiment with BraTS 2017 validation set shows that the proposed method achieves average dice scores of 0.757, 0.901, 0.825 for enhancing tumor core, whole tumor and tumor core, respectively.