AUTHOR=Liu Liangliang , Zhang Jing , Wang Jin-xiang , Xiong Shufeng , Zhang Hui 

TITLE=Co-optimization Learning Network for MRI Segmentation of Ischemic Penumbra Tissues

JOURNAL=Frontiers in Neuroinformatics

VOLUME=Volume 15 - 2021

YEAR=2021

URL=https://www.frontiersin.org/journals/neuroinformatics/articles/10.3389/fninf.2021.782262

DOI=10.3389/fninf.2021.782262

ISSN=1662-5196

ABSTRACT=Convolutional neural networks (CNNs) show promise for use in auxiliary diagnosis based on medical images. However, the scarcity of labeled medical image data represents a bottleneck that limits the performance of supervised CNN methods. In addition, annotating large amounts of labeled medical image data is often expensive and time-consuming. In this study, we propose a co-optimization learning network (COL-Net) for the segmentation of magnetic resonance images of ischemic penumbra tissues.
COL-Net is suitable for use with a limited number of labeled samples and consists of an unsupervised reconstruction network, a supervised segmentation network, and a transfer block. The reconstruction network extracts robust features by reconstructing pseudo-unlabeled samples; this represents the auxiliary branch of the segmentation network. The segmentation network is used to segment target lesions based on limited labeled samples and the auxiliary of the %AQ:  Please check to the phrase "and the auxiliary of the reconstruction network" as the meaning does not seem clear. reconstruction network. The transfer block is used to co-optimization feature maps between the bottlenecks of the reconstruction and segmentation networks. We propose a mixed loss function to optimize COL-Net. COL-Net is verified using a public ischemic penumbra segmentation challenge with 24 labeled samples. The results demonstrate that COL-Net has high predictive accuracy and generalization ability, with a Dice coefficient of 0.79. Further experiments show that COL-Net outperforms most supervised segmentation methods. COL-Net is thus a meaningful attempt to alleviate the limited labeled sample problem in medical image segmentation.