Abstract: Utilizing the powerful feature learning ability of deep learning, researchers have proposed a variety of effective methods for brain tumor segmentation in three-dimensional medical images. However, the existing approaches often need to be processed in stages pipeline, without considering the anatomical nested structural characteristics of brain tumors, thus affecting the accuracy and efficiency of tumor segmentation. In this paper, we propose the Nested Multi-Attention mechanism Network (NMA-Net) tailored for brain tumors, which can effectively connect the sub-segmentation tasks of different organizations, and can directly conduct end-to-end training. By using the segmentation result of the tumor peripheral tissue as a kind of soft attention to the tumor segmentation task, it can make the subsequent network focus more on the region of interest, which makes it possible to obtain more accurate segmentation results. Besides, we transform multi-class segmentation tasks into multiple binary sub-segmentation tasks. Experiments on the BraTS'2017 Challenge Dataset show that the proposed NMANet framework is very suitable for organ tissue segmentation with nested anatomical structures. Here, our single-view model achieves the best segmentation performance compared with the exiting approaches, and the multi-view fusion model also achieves the state-of-the-art performance on the TC and ET sub-regions.
Authors: Xinrui Zhuang (Tsinghua University, China); Yujiu Yang (Tsinghua University, Shenzhen, China)
Email: zhuangxr17@mails.tsinghua.edu.cn, yang.yujiu@sz.tsinghua.edu.cn