基于VMF-UNet的液基细胞制染机缺陷图像分割  

VMF⁃UNet based image segmentation of defects in liquid⁃based cell preparation staining machine

作　　者：田文豪 汪繁荣[1] 乔一航 TIAN Wenhao;WANG Fanrong;QIAO Yihang(School of Electrical and Electronic Engineering,Hubei University of Technology,Wuhan 430074,China)

机构地区：[1]湖北工业大学电气与电子工程学院,湖北武汉430074

出　　处：《现代电子技术》2025年第5期36-42,共7页Modern Electronics Technique

基　　金：国家自然科学基金项目(61903129)。

摘　　要：为了解决液基细胞制染机的成片效果缺陷识别问题,补偿设备最终成片率,文中提出一种VMF-UNet模型,模型以UNet为基础模型,使用VGG16Net的卷积部分替代UNet神经网络的编码器部分,加入多尺度高效局部注意力机制(MELA),引入特征细化模块(FRM),解决了图像过分割、欠分割、成片缺陷区域边缘不明显、UNet模型视野受限的问题。实验以医学检验可用性为原则,将显微镜下分割的数据集标准标签作为“金标准”。基于自建液基细胞成片缺陷区图像数据集的实验表明,改进网络在分割时平均交并比(MIo U)、平均像素精确度(MPA)、F1-score与准确率(Accuracy)分别为:82.73%、93.56%、81.93%、96.10%。实验结果证明,VMF-Unet模型对液基细胞制染机成片缺陷区域分割效果更好,可以有效补偿液基细胞制染机的最终成片率,为设备复处理提供有效依据,提高液基细胞制染机的广泛可应用性。A VMF-UNet model is proposed to identify the defects in the preparation effect of liquid-based cell preparation staining machine and compensate for its final preparation rate.This model is based on the UNet.The convolutional part of VGG16Net is used to replace the encoder part of UNet neural network.A multi-scale efficient local attention(MELA)mechanism is added and a feature refinement module(FRM)is introduced,which aims to eliminate the image over-segmentation,image under-segmentation,unclear edges of preparation defect areas,and limited field of view of the UNet model.The experiments are based on the principle of medical testing availability.The standard labels of the dataset segmented by the microscope are used as the″golden standard″.The experiments based on the self-built liquid-based cell preparation defect area image dataset show that the improved network has a mean intersection over union(MIoU),mean pixel accuracy(MPA),F1-score,and accuracy rate of 82.73%,93.56%,81.93%and 96.10%during segmentation.The experimental results demonstrate that the VMF-UNet model has a better segmentation effect on the preparation defect areas of the liquid-based cell preparation staining machine,and can effectively compensate for the final preparation rate of the machine,providing effective basis for equipment reprocessing and improving its applicability.

关 键 词：深度学习 语义分割 UNet 注意力机制 缺陷检测 液基细胞制染机 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TP391.4[电子电信—信息与通信工程]