Abstract: There are few algorithms for segmenting cryosection brain images, and most existing segmentation techniques presented limited precision and low efficiency. To address these problems, this paper proposed a novel deep feature learning-based segmentation algorithm using sparse autoencoder（SAE）. At the stage of feature extraction, SAE is trained twice to enhance the discriminability of the deep-learned feature representations. At the stage of classification, a softmax classifier is used for segmenting different objects. Experimental results of white matter segmentation on the Chinese Visible Human（CVH） dataset and its 3-D reconstruction show that, the learned deep feature performs much better in discriminability compared with other representative hand-crafted features（such as intensity, histogram of oriented gradient and principal components analysis） and achieves higher recognition accuracy.