LAUSR

Abstract As manual delineation of lesions in medical image is a very tedious and time consuming process, accurate and automatic segmentation of medical images can assist diagnosis and treatment. In this study, we propose a deep convolution neural network for stroke magnetic resonance imaging(MRI) segmentation. The main structure of our network consists of two symmetrical deep sub-networks, in which dense blocks are embedded for extracting effective features from sparse pixels to alleviate the over-fitting problem of deep networks. We use the multi-kernel to divide the network into two sub-networks for acquiring more receptive fields, and the dropout regularization method to achieve an effective feature mapping. For the post-processing of the soft segmentation, we use image median filtering to alleviate noises and preserve the edge details of images. Our network is evaluated on two public benchmark segmentation challenges (SISS: sub-acute ischemic stroke lesion segmentation and SPES: acute stroke outcome/penumbra estimation) with multi-modality MRI sequences. According to the results of the public benchmark reports, among 9 teams participating in both SISS and SPES challenges at the same time, our network achieves the top performance on SISS challenge, and the top 3 performance on the SPES challenge. In addition, our network also exhibits state-of-the-art performance compared with other segmentation methods. Finally, we extensively evaluate our network with an ablation experiment. The experimental results show that both multi-kernel and dropout strategies can improve the segmentation accuracy of our proposed network.