基于Transformer结构的深度学习模型用于外周血白细胞检测  

Deep learning model based on Transformer architecture for peripheral blood leukocyte detection

作　　者：冷冰[1,2] 冷敏 常智敏 葛明锋[1,2] 董文飞 Leng Bing;Leng Min;Chang Zhimin;Ge Mingfeng;Dong Wenfei(School of Biomedical Engineering(Suzhou),Division of Life Sciences and Medicine,University of Science and Technology of China,Suzhou 215163,China;Suzhou Institute of Biomedical Engineering and Technology,Chinese Academy of Science,Suzhou 215163,China;Liaocheng Cardiac Hospital,Liaocheng 252200,China)

机构地区：[1]中国科学技术大学生物医学工程学院(苏州),生命科学与医学部,苏州215163 [2]中国科学院苏州生物医学工程技术研究所,苏州215163 [3]聊城市心脏病医院,聊城252200

出　　处：《仪器仪表学报》2023年第5期113-120,共8页Chinese Journal of Scientific Instrument

基　　金：国家重点研发计划(2021YFB3602200)项目资助。

摘　　要：虽然血细胞分析仪已广泛应用于医院中,但人工镜检仍是白细胞检测的“金标准”。本文提出了一种基于DETR的Transformer结构的深度学习模型T-DETR用于外周血白细胞的检测,旨在缓解人工镜检的压力。首先,使用PVTv2作为DETR的骨干提取多尺度特征图来提高检测精度。然后,将可变形注意力模块引入到DETR模型中,减少计算复杂度以加快模型收敛。最后,为了得到最优权重,在筛选后的公共白细胞数据集上使用了迁移学习的训练方式。实验结果表明,T-DETR在COCO数据集上mAP为0.476,在白细胞数据集上的mAP为0.954,优于DETR和经典CNN模型,验证了Transformer结构的模型在医学图像检测中应用的可行性。Although blood cell analyzers have been widely used in hospitals,the manual microscopy is still the “gold standard” for leukocyte detection.In this article,T-DETR,a DETR-based deep learning model with Transformer architecture is proposed for the detection of peripheral blood leukocytes,which aims to relieve the pressure of manual microscopy.First,PVTv2 is used as the backbone of DETR to extract multi-scale feature maps to improve detection accuracy.Then,the deformable attention module is introduced into the DETR model to reduce the computational complexity to accelerate the model convergence.Finally,to obtain the optimal weights,the training method of transfer learning is used on the filtered public leukocyte dataset.Experimental results show that T-DETR has an mAP of 0.476 on the COCO dataset and 0.954 on the leukocyte dataset,which is better than DETR and the classical CNN model.Results verify the feasibility of the Transformer structured model for medical image detection applications.

关 键 词：白细胞 目标检测 深度学习 TRANSFORMER DETR 

分 类 号：TH776[机械工程—仪器科学与技术]