A framework combining DNN and level-set method to segment brain tumor in multi-modalities MR image

Abstract

The magnetic resonance (MR) brain tumor image segmentation can quantitatively analyze the tumor size and provide a large number of brain functional and anatomical information, which to a certain degree can guide the brain disease diagnosis and treatment planning. In this paper, we proposed a framework for brain tumor MR image segmentation combining deep learning and level-set method. First of all, we trained deep neural network (DNN) to classify center pixel of the image patches according to four MR modalities (T1, T1c, T2 and flair) and generated the segmentation result as initialization for level-set method. Secondly, we refine the segmentation results in edema region by level set in flair modality, which compensated for the discontinuity of the DNN segmentation improving the segmentation accuracy. In order to balance tumor patches proportion in datasets increasing them from 2 to 15%, we select patches randomly with a fixed proportion between tumor and normal tissue. Experiments show that the proposed method can effectively overcome discontinuity in segmentation result and obtain a satisfied segmentation results.