西安电子科技大学学报 ›› 2022, Vol. 49 ›› Issue (2): 198-206.doi: 10.19665/j.issn1001-2400.2022.02.023

• 计算机科学与技术 & 网络空间安全 • 上一篇    下一篇

注意力驱动的细胞团簇细胞核分割算法

马思珂1(),赵萌1(),石凡1(),孙续国2(),陈胜勇1()   

  1. 1.天津理工大学 计算机视觉与系统教育部重点实验室,天津 300384
    2.天津医科大学 医学检验学院,天津 300203
  • 收稿日期:2020-09-06 出版日期:2022-04-20 发布日期:2022-05-31
  • 通讯作者: 赵萌
  • 作者简介:马思珂(1996—),女,天津理工大学硕士研究生,E-mail: sike_ma@126.com;|石 凡(1984—),男,副教授,博士,E-mail: shifan@email.tjut.edu.cn;|孙续国(1965—),男,教授,博士,E-mail: sunxuguo@tmu.edu.cn;|陈胜勇(1973—),男,教授,博士,E-mail: sy@ieee.org
  • 基金资助:
    国家自然科学基金(61703304);国家自然科学基金(61906133);国家自然科学基金(U1509207);广东省重点领域研发计划(2019B010109001);天津市重大科研项目(18ZXZNGX00150)

Attention driven nuclei segmentation method for cell clusters

MA Sike1(),ZHAO Meng1(),SHI Fan1(),SUN Xuguo2(),CHEN Shengyong1()   

  1. 1. Key Laboratory of Computer Vision and System of Ministry of Education, Tianjin University of Technology,Tianjin 300384
    2. School of Medical Laboratory,Tianjin Medical University,Tianjin 300203
  • Received:2020-09-06 Online:2022-04-20 Published:2022-05-31
  • Contact: Meng ZHAO

摘要:

胸腔积液细胞团簇的细胞核形态为肺癌诊断、肿瘤转移及治疗效果评价提供了重要途径,而对其细胞核进行精准分割是肺癌病理诊断工作的基础。由于胸腔积液肿瘤细胞团簇复杂的生成背景,以及细胞核特征的不均匀性(特征信息分散)和团簇内部重叠细胞中的触核情况(特征不明显),使细胞团簇分割仍然是一个具有挑战性的问题。提出了基于注意力机制的改进U-Net模型,从空间注意力和通道注意力两方面来增强对细胞核非显著特征的学习;并改进U-Net的跳跃连接,融合U-Net中深层和浅层特征,解决语义间隙的问题。实验结果表明,与最新的其他方法相比,CRUNet能够在所建立的胸水细胞团簇数据集上取得更好的分割效果。为了进一步说明该网络的有效性,在公共数据集BBBC020上也与其他网络进行了对比。

关键词: 胸腔积液细胞团簇, 图像分割, 注意力机制, U-Net

Abstract:

Nuclei morphology of pleural effusion cell clusters provides an essential way for the diagnosis,metastasis,and treatment evaluation of the lung cancer.Accurate segmentation of the nuclei is the basis of pathological diagnosis of the lung cancer.Because of the complex background of tumor cell clusters in pleural effusion,the inhomogeneity of nuclei features (scattered feature information),and nuclei overlapping within clusters (whose characteristics are not prominent),the segmentation of tumor cell clusters is still a challenging problem.In this paper,an improved U-Net model,named CRUNet,based on the attention mechanism is proposed.With the attention module,the CRUNet can enhance the learning of non-salient features of the nucleus from spatial attention and channel attention,and improve the jumping connection of the U-Net to integrate the deep and shallow features of the U-Net to solve the problem of the semantic gap.Experimental results show that compared with other state-of-the-art methods,the CRUNet can achieve a better segmentation performance on our self-established pleural effusion cell cluster dataset.To further illustrate the effectiveness of the proposed network,the CRUNet is also compared with other networks on a public cell dataset-BBBC020.

Key words: pleural effusion cell clusters, segmentation, attention mechanism, U-Net

中图分类号: 

  • TP391