基于语义分割及主动轮廓的宫颈细胞实例分割方法  

Instance segmentation of cervical cell based on semantic segmentation and active contour

作　　者：朱琳琳 杜泓 ZHU Lin-lin;DU Hong(College of Automation,Shenyang Aerospace University,Shenyang 110136,China)

机构地区：[1]沈阳航空航天大学自动化学院,沈阳110136

出　　处：《沈阳航空航天大学学报》2022年第2期55-63,共9页Journal of Shenyang Aerospace University

基　　金：国家自然科学基金(项目编号:61873259,61903262),辽宁省自然基金(项目编号:2018551107)。

摘　　要：宫颈细胞图像的分割问题研究对宫颈癌筛查诊断具有重要意义,但大量的重叠细胞给这项研究带来了巨大挑战。针对这一问题提出了新的宫颈细胞实例分割方法,该方法在语义分割的基础上,根据先验信息为细胞初始化轮廓并建立能量函数,通过最小化能量函数完成轮廓演化,解决复杂宫颈细胞图像中细胞实例分割问题。首先,改进U-net神经网络,完成宫颈细胞图像精确语义分割,将图像像素分为细胞、细胞核及背景3类;其次,在语义分割的基础上,识别图像中的游离细胞与细胞团块,并直接提取游离细胞的轮廓;最后,对于包含重叠细胞的细胞团块,使用Snake主动轮廓模型完成细胞团块中每个细胞的轮廓提取,实现细胞的实例分割。实验结果表明,算法可对包含重叠细胞的宫颈细胞图像进行有效实例分割。Cervical cell segmentation is import to the diagnosis of cervical cancer,in which a large number of overlapping cells have brought great challenges.A new cervical cell instance segmentation method to solve this problem was proposed.This method initialized the contour for each cell based on the result of semantic segmentation and establishes an energy function for each contour.The contour evolution was completed by minimizing the energy function,which solved the problem of segmentation of overlapping cells in TCT images.First,the U-net neural network to complete the precise semantic segmentation of cervical cell images was improved,and the image pixels were divided into three categories:cells,nuclei,and background.Second,on the basis of semantic segmentation,free lying cells and cell clumps in the image were identified and directly extract the contours of the free lying cells.Last,for the cell clumps containing overlapping cells,the snake active contour model was used to extract the contour of each cell to achieve cell instance segmentation.The experimental results show that the proposed algorithm can perform effective instance segmentation on cervical cell images containing overlapping cells.

关 键 词：宫颈细胞图像 U-net网络 主动轮廓 重叠细胞 实例分割 语义分割 

分 类 号：O235[理学—运筹学与控制论]