基于对比学习和轮廓引导U-Net的细胞核分割-《中国医学物理学杂志》

文章信息/Info

Title:: Nuclei segmentation using contrastive learning and contour guided U-Net

作者:: 贾伟1; 江海峰2; 赵雪芬3; 1.宁夏大学信息工程学院，宁夏银川 750021； 2.宁夏医科大学总医院病理科，宁夏银川 750021； 3.宁夏大学新华学院，宁夏银川 750021

Author(s):: JIA Wei1; JIANG Haifeng2; ZHAO Xuefen3; 1. School of Information Engineering, Ningxia University, Yinchuan 750021, China 2. Department of Pathology, General Hospital of Ningxia Medical University, Yinchuan 750021, China 3. Xinhua College, Ningxia University, Yinchuan 750021, China

Keywords:: Keywords: contrastive learning contour feature U-Net nuclei segmentation active contour

摘要:: 针对细胞核分割中遇到的细胞核边缘模糊和重叠问题，提出一种基于对比学习和轮廓引导U-Net的细胞核分割方法。首先，为解决已标记数据不足的问题，提出全局和局部特征对比学习，利用未标记病理图像的全局特征和局部特征进行对比学习。然后，为提高U-Net的分割性能，提出轮廓引导的U-Net对每层编码器的轮廓特征进行融合，并利用融合后的轮廓特征和Jeffreys散度驱动的主动轮廓方法引导和辅助细胞核分割。最后，利用全局和局部特征对比学习与轮廓引导的U-Net进行交替训练，将置信度较高的未标记数据作为伪标签，混入到已标记数据中，再使用轮廓引导的U-Net进行分割训练。实验结果表明，该分割方法能够有效提高细胞核分割的准确性，具有较为稳定的分割性能。

Abstract:: Abstract: Aiming at the problems of nucleus edge blurring and overlap in nucleus segmentation, a nuclei segmentation method based on contrastive learning and contour guided U-Net is proposed. The global and local feature contrastive learning is adopted for contrastive learning using the global and local features of unlabeled pathological images, thereby solving the problem of insufficient labeled data. Then a contour guided U-Net is developed to fuse the contour features of each layer of encoder for improving the segmentation performance of U-Net, and the fused contour features and Jeffreys divergence-driven active contour method are used to guide and assist the nuclei segmentation. Finally, the alternate training using the global and local feature contrastive learning and the contour guided U-Net is carried out. The unlabeled data with high confidence are taken as pseudo-label and mixed into labeled data, and the contour guided U-Net is used for segmentation training. Experimental results show that the proposed segmentation method can effectively improve the accuracy of nuclei segmentation, and has a relatively stable segmentation performance.