轻量化深度网络辅助于无透镜计算显微图像的细胞分类  

Lightweight Deep Learning Network Assisted Cell Classification Using Lensless Computational Microscopic Imaging Data

作　　者：王朝晖 康欢 陈多芳 徐欣怡 曾琦 梁继民[2] 陈雪利 Wang Zhaohui;Kang Huan;Chen Duofang;Xu Xinyi;Zeng Qi;Liang Jimin;Chen Xueli(Xi’an Key Laboratory of Intelligent Sensing and Regulation of Trans-Scale Life Information,School of Life Science and Technology,Xidian University,Xi'an 710126,Shaanxi,China;School of Electronic Engineering,Xidian University,Xi'an 710126,Shaanxi,China)

机构地区：[1]西安电子科技大学生命科学技术学院西安市跨尺度生命信息智能感知与调控重点实验室,陕西西安710126 [2]西安电子科技大学电子工程学院,陕西西安710126

出　　处：《中国激光》2022年第5期130-138,共9页Chinese Journal of Lasers

基　　金：中央高校基本科研业务费(QTZX2185,JB211211);国家重点研发计划(2018YFC0910600);国家自然科学基金(81871397,61901338);“万人计划”人才、陕西省杰出青年科学基金(2020JC-27);陕西省“特支计划”青年拔尖人才。

摘　　要：无透镜计算显微成像是一种低成本、高效的成像技术。这种成像方式具有大视野、高通量的特点,能够实时地对细胞进行无标记成像。提出了一种轻量化网络模型(Depthwise-ResNeXt),将该神经网络与无透镜计算显微成像进行有机结合,实现了实时准确的细胞分类。使用SUM、MCF10A、ECa109、CL-1四种细胞作为分类数据,Depthwise-ResNeXt对这四类细胞的分类准确率达到92.8%,参数量仅有806kB。该网络证明了神经网络与无透镜计算显微成像在细胞分类领域相结合的可能性,并大大降低了神经网络在细胞分类方面的应用成本。Objective Cell analysis is one of the most important application scenarios of microscopic images, and it plays a vital role in biological research, clinical medicine, drug screening, and other fields. With the rise of artificial intelligence, neural networks have become an important method for cell classification. The neural network has a complex structure, many parameters, a large amount of calculations, and requires advanced graphics cards for training. Taking account of the computer performance of hospitals and biological research institutions, as well as the popularity of mobile terminals today, it is necessary to reduce the computational overhead of neural networks. It can reduce the corresponding medical and scientific research costs, which is conducive to the application and promotion of neural networks for cell classification. It is an urgent need for a method to accurately classify cells in real time and at low cost.Methods Lensless computational microscopic imaging is a low-cost and efficient imaging technology. This imaging method has the characteristics of large field-of-view and high throughput. It can image cells in real time without mechanical focusing or staining. This paper proposes a lightweight network model named Depthwise-Res Ne Xt to accurately classify cells in real time through the seamless integration of the neural network and lensless computational microscopy. The imaging system(Fig. 1) consists of a laser source, micro-holes and sensors. The laser emitted from the light source passes through the micro-hole and the sample on the holder, and irradiates the sensor. The sensor collects the light signal and transmits it to the computer. The system uses a laser diode as the light source with an output power of 1.5 m Wand a wavelength of 532 nm. The diameter of micro-holes is 40 μm. The signal acquisition uses a CMOS sensor with a pixel size of 1.67 μm. The distance between the sensor and the sample is within 1 mm.We collect the data of four types of cells including SUM, MCF10A, ECa109, and

关 键 词：生物光学 数字全息 计算显微成像 无透镜成像技术 细胞分类 神经网络 轻量化网络 

分 类 号：TN26[电子电信—物理电子学] TP181[自动化与计算机技术—控制理论与控制工程]