Pre-training can alleviate the requirement of labeling data for a new task. However, Pre-training as a sequential learning typically suffers in fact from forgetting the older tasks. Especially in complex medical image segmentation tasks, this problem is more prominent. To solve above problem, we propose a network structure based on feature space transformation (FS-Net) for data expansion of medical image segmentation. FS-Net share parameters during training to help exploiting regularities present across tasks and improving the performance by constraining the learned representation. In the experiment, we use M&Ms as the extended dataset of HVSMR, these two tasks have the same segmentation target (heart). The segmentation accuracy of FS-Net is up to 7.12% higher than the baseline network, which is significantly better than Pre-training. In addition, we use Brats2019 as expansion dataset on WMH, and the segmentation accuracy is improved by 0.77% compared with the baseline network. And Brats2019 (glioma) and WMH (white matter hyperintensities) have different segmentation targets.
基金：
National Key Research and Development Program of China [2018YFC1312000]; Basic Research Foundation of Shenzhen Science and Technology Stable Support Program [GXWD20201230155 427003-20200822115709001]
语种：
外文
WOS：
第一作者：
第一作者机构：[1]Harbin Inst Technol Shenzhen, Dept Elect & Informat Engn, Shenzhen, Peoples R China
通讯作者：
通讯机构：[1]Harbin Inst Technol Shenzhen, Dept Elect & Informat Engn, Shenzhen, Peoples R China[2]Peng Cheng Lab, Shenzhen, Guangdong, Peoples R China[3]Capital Med Univ, Xuanwu Hosp, Natl Clin Res Ctr Geriatr Disorders, Beijing, Peoples R China[4]Capital Med Univ, Adv Innovat Ctr Human Brain Protect, Beijing, Peoples R China
推荐引用方式(GB/T 7714)：
Guo Xutao,Yang Yanwu,Ma Ting.FS-Net: A New Paradigm of Data Expansion for Medical Image Segmentation[J].DEEP GENERATIVE MODELS, AND DATA AUGMENTATION, LABELLING, AND IMPERFECTIONS.2021,13003:217-225.doi:10.1007/978-3-030-88210-5_21.
APA：
Guo, Xutao,Yang, Yanwu&Ma, Ting.(2021).FS-Net: A New Paradigm of Data Expansion for Medical Image Segmentation.DEEP GENERATIVE MODELS, AND DATA AUGMENTATION, LABELLING, AND IMPERFECTIONS,13003,
MLA：
Guo, Xutao,et al."FS-Net: A New Paradigm of Data Expansion for Medical Image Segmentation".DEEP GENERATIVE MODELS, AND DATA AUGMENTATION, LABELLING, AND IMPERFECTIONS 13003.(2021):217-225
