基于昇腾NPU的癌细胞检测算法设计及实现  

Design and Implementation of Cancer Cell Detection Algorithm Based on NPU*

作　　者：吴振宁 肖仲喆 江均均 黄敏 WU Zhenning;XIAO Zhongzhe;JIANG Junjun;HUANG Min(School of Optoelectronic Science and Engineering,Soochow University,Suzhou Jiangsu 215006,China)

机构地区：[1]苏州大学光电科学与工程学院,江苏苏州215006

出　　处：《电子器件》2020年第6期1210-1214,共5页Chinese Journal of Electron Devices

基　　金：国家自然科学基金项目(61906128,61802272);江苏省自然科学基金项目(BK20180834)。

摘　　要：为了解决临床针图像中循环肿瘤细胞(CTC)人工检测效率低的问题,提出了利用经典图像处理方法进行预处理并利用卷积神经网络(CNN)进行判断识别的解决方法。通过预处理初步检测出图像中所有的疑似癌细胞,将得到的细胞图像输入到训练好的网络进行判断并得到检测结果。实验中采用临床采集的图像进行测试,测试过程中网络判别准确率为90%,且没有出现漏判。结果表明:利用卷积神经网络的癌细胞识别方法具有可靠的效果,相较于人工判断具有精度上的优势,能够作为癌细胞识别的重要手段。同时,算法在集成了NPU加速芯片的华为Atlas200 DK硬件平台上运行,实现了运算加速,并为实现应用的离线部署创造了条件。It’s inefficient for human to detect the circulating tumor cells(CTC)in clinical needle images,In order to solve the problem,this paper proposes a solution using classic image processing methods for preprocessing and convolutional neural network(CNN)for judgment.The pre-processing is used to check all the suspected cancer cells in the image,the obtained cell image will be input to the network.After a few seconds,it can give us the detection result.In the experiment,clinically collected images were used for testing.During the test,the accuracy of the network is 90%and there was no omission.The results show that the method of cancer cell recognition using convolutional neural network has a reliable effect.Compared with manual judgment,it has an advantage in accuracy,so it can be used as an important method for cancer cell recognition.At the same time,the algorithm runs on the Huawei Atlas200 DK hardware platform with an integrated NPU acceleration chip,which accelerates computation and creates conditions for offline deployment of applications.

关 键 词：癌细胞检测 预处理 NPU 卷积神经网络 

分 类 号：R730.4[医药卫生—肿瘤]