Abstract
Accurate segmentation of MR brain tissue is a crucial step fordiagnosis,surgical planning, and treatment of brain abnormalities. However,itis a time-consuming task to be performed by medical experts. So, automatic andreliable segmentation methods are required. How to choose appropriate trainingdataset from limited labeled dataset rather than the whole also has greatsignificance in saving training time. In addition, medical data labeled is toorare and expensive to obtain extensively, so choosing appropriate unlabeleddataset instead of all the datasets to annotate, which can attain at least sameperformance, is also very meaningful. To solve the problem above, we design anautomatic segmentation method based on U-shaped deep convolutional network andobtain excellent result with average DSC metric of 0.8610, 0.9131, 0.9003 forCerebrospinal Fluid (CSF), Gray Matter (GM) and White Matter (WM) respectivelyon the well-known IBSR18 dataset. We use bootstrapping algorithm for selectingthe most effective training data and get more state-of-the-art segmentationperformance by using only 50% of training data. Moreover, we propose a strategyof MR brain tissue images' suggestive annotation for unlabeled medical databased on the modified U-net. The proposed method performs fast and can be usedin clinical.