基于深度迁移学习的前列腺癌细胞分类  

Prostate Cancer Cell Classification Based on Deep Transfer Learning

作　　者：唐李文 曹依琳 汪雅莉 平斯羽 胡文静 赵林[1,2] TANG Liwen;CAO Yilin;WANG Yali;PING Siyu;HU Wenjing;ZHAO Lin(School of Information Science and Engineering,Hunan Institute of Science and Technology,Yueyang 414006,China;Research Center of Machine Vision and Artificial Intelligence,Hunan Institute of Science and Technology,Yueyang 414006,China)

机构地区：[1]湖南理工学院信息科学与工程学院,湖南岳阳414006 [2]湖南理工学院机器视觉及人工智能研究中心,湖南岳阳414006

出　　处：《湖南理工学院学报（自然科学版）》2022年第2期7-12,共6页Journal of Hunan Institute of Science and Technology(Natural Sciences)

基　　金：湖南省自然科学基金项目(2019JJ40110,2020JJ4331);湖南省研究生科研创新项目(CX20201142);湖南省大学生创新创业训练计划项目(S202010543050)。

摘　　要：前列腺癌是我国发病率增长最快的癌症之一.高效准确的前列腺癌细胞分析方法研究是生物医学、临床诊断等领域的热点问题.目前,直肠指诊法、经直肠超声检查法、特异性抗原测定法等前列腺癌早期诊断方法需专业人员操作、荧光标记且耗时耗力.针对以上问题,采用正交偏振衍射成像流式细胞仪实验系统,使用相干线性激光束照射细胞,由显微物镜收集散射光并经相机采集获得前列腺细胞的衍射图像.由于前列腺细胞衍射图像较难获取,实验中使用的前列腺细胞数据集样本数量较少.提出一种基于迁移学习的细胞残差卷积神经网络的前列腺细胞分类方法.该免标记检测方法通过迁移学习可以在其他训练收敛的网络模型上进行微调,进而得到可有效分析前列腺细胞特征的参数,从而实现对前列腺细胞的免标记精准分类,分类准确率为96.19%.In recent years,prostate cancer has became the cancer with the fastest growing incidence in our country.The development of new methods for the analysis of prostate cancer cells with high efficiency and accuracy is a problem hot in the fields of biomedical science and clinical diagnosis.At present,early diagnosis methods for prostate cancer such as digital rectal examination,transrectal ultrasonography and specific antigen assay require professional operation,time-consuming,labor-intensive and fluorescent labeling.To solve the above problems,an orthogonal polarization diffraction imaging flow cytometer experimental system was used to irradiate the cells with a coherent linear laser beam in the paper,the scattered light was collected by a microscope objective,and the diffraction image of the prostate cells was collected by a camera.However,due to the difficulty of obtaining prostate cell diffraction images,the number of samples in the prostate cell data set used in the experiment was relatively small,so we proposed a new method of prostate cell classification based on cell residual convolutional neural network and transfer learning.The label-free detection method can be fine-tuned the parameter of trained convergent network through transfer learning,then we obtained the parameters that can analyze the characteristics of prostate cells effectively,so as to realize the label-free and accurate classification of prostate cells,and the classification accuracy rate is 96.19%.It is an effective method for early diagnosis of prostate cancer and has important biomedical and clinical diagnostic value.

关 键 词：前列腺癌 偏振成像 迁移学习 卷积神经网络 细胞免标记分类 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]