基于CycleGAN的虚拟定量差分相衬成像用于红细胞分类  

CycleGAN-Based Virtual Quantitative Differential Phase Contrast Imaging for Red Blood Cell Classification

作　　者：汪涛[1] 彭韬 蒋梦朵 张粲 张凯旋 陆凤雅 钟振声 周金华[1,2] WANG Tao;PENG Tao;JIANG Mengduo;ZHANG Can;ZHANG Kaixuan;LU Fengya;ZHONG Zhensheng;ZHOU Jinhua(School of Biomedical Engineering,Anhui Provincial Institute of Translational Medicine,Anhui Medical University,Hefei Anhui 230032,China;D-Printing and Tissue Engineering Center,Anhui Provincial Institute of Translational Medicine,Anhui Medical University,Hefei Anhui 230032,China)

机构地区：[1]安徽医科大学生物医学工程学院,安徽合肥230032 [2]安徽医科大学3D打印与组织工程中心、安徽省转化医学研究所,安徽合肥230032

出　　处：《中国医疗设备》2023年第4期1-6,12,共7页China Medical Devices

基　　金：安徽省重点研究与开发计划(2022a05020028);安徽省自然科学基金(2208085MC54);安徽省转化医学研究院科研基金(2021zhyx-B16);安徽高校自然科学研究重点项目(KJ2021A0251,2022AH050676)。

摘　　要：目的通过深度学习技术实现虚拟定量差分相衬(Virtual Quantitative Differential Phase Contrast,V-qDPC)重建,提高定量相位成像衬度和鲁棒性,为无标记红细胞的全自动分类提供新思路。方法通过对LED照明进行编码获得明场图像和差分相衬图像,通过相位重建可获得定量差分相衬(Quantitative Differential Phase Contrast,qDPC)图像;采用循环一致生成对抗网络(Cycle-consistent Generative Adversarial Network,CycleGAN)完成明场图像到qDPC图像的端到端映射。结果基于CycleGAN生成的V-qDPC图像,实验参数当λ=7和β=0.5时,V-qDPC图像质量最优;相比光学重建的qDPC图像有更好的鲁棒性和抗噪声能力;使用AlexNet、ResNet50和VggNet三种网络模型比较无标记红细胞形态的自动分类,结果表明V-qDPC图像比qDPC图像具有更好分类性能。结论与传统的基于多幅倾斜照明图像的qDPC重建相比,V-qDPC算法具有更好的相位图像质量和鲁棒性,适合以高精度和高效率实现全自动细胞分类,同时省去了成像光路和硬件支持,有望应用于生物医学研究。Objective Virtual quantitative differential phase contrast(V-qDPC)reconstruction is realized by deep learning technology,which improves the contrast and robustness of quantitative phase imaging,and provides a new idea for automatic classification of unlabeled red blood cells.Methods By encoding LED lighting,bright field(BF)images and differential phase contrast(DPC)images are obtained by programmable LED illumination,and quantitative differential phase contrast(qDPC)image can be obtained by phase reconstruction.End-to-end mapping of BF images to qDPC images is completed by cycle-consistent generative adversarial network(CycleGAN).Results Based on the V-qDPC image generated by CycleGAN,the quality of V-qDPC image was optimal when the experimental parameters wereλ=7 andβ=0.5.Compared with optical reconstruction qDPC image had better robustness and anti-noise ability;AlexNet,ResNet 50 and VggNet were used to compare the automatic classification of unlabeled red blood cell morphology.The results showed that V-qDPC images had better classification performance than qDPC images.Conclusion Compared with the traditional qDPC reconstruction based on several oblique lighting images,V-qDPC algorithm has better phase image quality and robustness.It is suitable for realizing automatic cell classification with high precision and high efficiency,while eliminating the imaging optical path and hardware support.It is expected to be applied in biomedical research.

关 键 词：定量差分相衬成像 无标记红细胞 循环一致生成对抗网络 全自动分类 

分 类 号：R197.39[医药卫生—卫生事业管理]