基于轻量网络的白细胞快速检测算法  

Fast White Blood Cell Detection Algorithm Based on Lightweight Network

作　　者：陈亮 郭慧慧 尹涛 Chen Liang;Guo Huihui;Yin Tao(School of Information and Electrical Engineering,Hunan University of Science and Technology,Xiangtan 411201,Hunan,China)

机构地区：[1]湖南科技大学信息与电气工程学院,湖南湘潭411201

出　　处：《中国生物医学工程学报》2025年第1期66-76,共11页Chinese Journal of Biomedical Engineering

基　　金：国家自然科学基金(62271199);湖南省研究生科研创新项目(CX20221054)。

摘　　要：白细胞由于种类繁多、形态差异性大,且在血液镜检时常存在重叠、粘连、细胞边界模糊及色变等问题,传统基于图像检测的系统特征提取困难,检测精度较差,稳定性不足。针对上述问题,本研究提出一种基于轻量网络结构的白细胞快速检测算法。首先采用MobileNetv3为特征提取网络,并提出了一种双通道金字塔特征融合结构TCPF-Net完成特征融合,提升了算法对模糊、色变、形态各异的白细胞图像特征提取能力;然后,针对白细胞特殊的长宽比与尺度特征,舍弃检测网络的大目标检测头而仅保留中小目标检测头,提升算法对白细胞的检测速度;最后,采用完整锚框与目标重叠时的交并比参数对检测网络位置回归损失函数进行优化,提升算法对重叠、粘连细胞的检测能力。采用基于瑞氏染色法染色后的人体血液40倍显微图像进行实验,通过对8848张白细胞图像进行实验验证,该轻量网络算法白细胞检测平均精度均值(mAP)达98.8%,较改进前网络提高1.1%,同时每秒处理图像的帧数(FPS)达54.19,较改进前提升了32%,实现了白细胞快速而精准的检测。Due to the large variety and morphological differences of white blood cells,and overlap,adhesion,cell boundary blurring and color change in blood microscopy,the traditional system based on image detection has difficulty in feature extraction,poor detection accuracy and insufficient stability.To address these problems,a white blood cell rapid detection algorithm based on lightweight network structure was proposed.Firstly,the algorithm used MobileNetv3 as the feature extraction network,and proposed a dual-channel pyramid feature fusion structure TCPF-Net to complete the feature fusion for its insufficient feature extraction ability.The algorithm improved the feature extraction ability of white blood cell images with blur,color change and different shapes.After that,the algorithm abandoned the large target detection head of the detection network and only retained the small and medium target detection head for the special aspect ratio and scale characteristics of white blood cells,which improved the detection speed of the algorithm for the white blood cells.Finally,the algorithm used the intersection over union parameter when the complete anchor frame overlaped with the target to complete the optimization of the regression loss function of the detection network position,and improved the detection ability of the algorithm for overlapping and adherent cells.The experiment was conducted using 40x microscopic images of human blood stained with the Romanowsky staining method.With the validation of 8,848 white blood cell images,the meanaverage precision(mAP)of the lightweight network algorithm for white blood cell detection reached 98.8%,representing 1.1%improvement compared to the original network.Simultaneously,theframes per second(FPS)reached 54.19,indicating a 32%increase compared to the original network,achieving the rapid and precise detection of white blood cells.

关 键 词：白细胞检测 深度学习 双通道特征融合 轻量化网络 

分 类 号：R318[医药卫生—生物医学工程]