结合通道空间加权特征金字塔网络的白细胞检测与分割  

White Blood Cell Detection and Segmentation Combined with the Channel Space Weighted Feature Pyramid Networks

作　　者：耿磊[1,2] 杨盟盟 肖志涛 张芳[1,2] 刘彦北 吴骏[2,3] 王雯 Geng Lei;Yang Mengmeng;Xiao Zhitao;Zhang Fang;Liu Yanbei;Wu Jun;Wang Wen(School of Life Sciences,Tiangong University,Tianjin 300387;Tianjin Key Laboratory of Optoelectronic Detection Technology and System,Tianjin 300387;School of Electronics and Information Engineering,Tiangong University,Tianjin 300387)

机构地区：[1]天津工业大学生命科学学院,天津300387 [2]天津市光电检测技术与系统重点实验室,天津300387 [3]天津工业大学电子与信息工程学院,天津300387

出　　处：《计算机辅助设计与图形学学报》2021年第9期1418-1427,共10页Journal of Computer-Aided Design & Computer Graphics

基　　金：天津市科技重大专项与工程(17ZXSCSY00060,17ZXHLSY00040).

摘　　要：白细胞的准确检测和精确分割是一项具有挑战性的医学图像处理任务.在显微镜下获取的白细胞图像会受到染色杂质的影响,且白细胞种类繁多、形态各异、类间差别小,还存在相互重叠相互粘连的现象,导致细胞边缘无法被准确分割,上述问题一直都是白细胞图像检测和分割的难点.针对以上问题,基于Mask R-CNN提出了结合注意力机制多尺度特征融合的白细胞检测方法.在Mask R-CNN结构的基础上,在特征金字塔网络(FPN)模块中融合了注意力机制模块,提出了通道空间加权特征金字塔网络.该结构不仅可以学习特征图中重要通道特征的权重大小,还可以学习层中重要特征区域的表示.同时,在网络结构中加入了Skip-FPN模块,该模块通过短连接融合更多白细胞的底层细节信息,从而更准确地检测白细胞,更精确地进行白细胞的形态分割.实验结果表明,所提方法具有良好的检测与分割性能.在Kaggle开源数据集下,所提方法对白细胞检测的指标mAP值达到了98.25%,与改进前相比提高了1.25%;分割的平均精度mIoU值达到了89.30%,与改进前相比提高了0.002%.The accurate detection and segmentation of leukocytes is a challenging task in medical image processing.The white blood cell image obtained under microscope is easily affected by impurities.There are many kinds of white cells,different shapes and small differences among them,and also exiting overlapping and adhesion phenomena,which leads to the inaccuracy of cell edge segmentation.The above problems are always the difficulties of detection and segmentation of white cell image.To solve the above problems,a leukocyte detection method based on Mask R-CNN and attention mechanism multi-scale feature fusion is proposed.Based on the Mask R-CNN structure,proposed thesis integrates the attention mechanism module in the FPN(feature pyramid networks)module,and proposes the CSFPN(channel spatial feature pyramid networks)structure.This structure can learn the weight of important channel features and the representation of important feature regions in the feature maps.At the same time,Skip-FPN module is added to the network structure,which fuses more low-level detailed information of leukocytes through short connection,so as to detect and segment leukocytes more accurately.Experimental results show that this method has good detec-tion and segmentation performance.Under the Kaggle open source data set,the mAP value of this method for white blood cell detection reaches 98.25%,which has an increase of 1.25%compared with the previous improvement.The accuracy mIoU value reached 89.3%,which has an increase of 0.002%compared to the before improvement.

关 键 词：卷积神经网络 语义分割 多尺度特征 通道空间加权 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]