基于改进Mask R-CNN模型的宫颈细胞分割  

Cervical Cell Segmentation Based on Improvd Mask R-CNN Model

作　　者：李静 张悦 乔亚鑫 宁春玉[1] LI Jing;ZHANG Yue;QIAO Yaxin;NING Chunyu(School of Life Science and Technology,Changchun University of Science and Technology,Changchun 130022)

机构地区：[1]长春理工大学生命科学技术学院,长春130022

出　　处：《长春理工大学学报（自然科学版）》2024年第2期107-113,共7页Journal of Changchun University of Science and Technology(Natural Science Edition)

基　　金：吉林省科技发展计划项目(20220101123JC)。

摘　　要：在宫颈细胞分割过程中,原始Mask R-CNN模型采用ResNet50和FPN作为特征提取网络,尽管模型分割效果良好,但仍存在分割速度慢且边缘分割效果欠佳等问题,为此,提出了一种改进Mask R-CNN模型。首先,该模型采用轻量化网络MobileNet V2作为特征提取模块,大幅度降低模型参数量,为图像的实时分割提供了可能。其次,该模型在特征提取网络中融入了注意力模块,通过自适应特征优化功能,最大限度获取底层信息。最后,模型在掩码生成阶段采用跳跃连接的方式,有效融合各尺度信息,提升网络信息获取能力。实验结果表明,改进模型将宫颈细胞核的分割速度提升了50%左右、分割精度提升了7%。In the process of cervical cell segmentation,the original Mask R-CNN model utilizes ResNet50 and FPN as the feature extraction networks yielding satisfactory segmentation results.However,there still exist problems such as slow segmentation speed and suboptimal edge segmentation.To address these issues,an improved Mask R-CNN model is proposed.Firstly,the model employs the lightweight MobileNet V2 as the feature extraction module,which significantly reduces the amount of model parameters and provides the possibility of real-time segmentation for cervical cell images.Secondly,an attention module is integrated into the feature extraction network,which maximizes the access to the underlying information through the adaptive feature optimization function.Lastly,the model adopts the skip connection in the mask generation stage to effectively fuse the information of various scales,which improves the ability of acquiring information of the network.Experimental results demonstrate that the proposed model has increased the segmentation speed of cervical cell nuclei by about 50%and the segmentation accuracy by 7%.

关 键 词：细胞分割 深度学习 MaskR-CNN 注意力机制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]