基于CMOS双线阵图像传感器的片上细胞显微成像及处理芯片的研究

The collection and detection of cell image in biological samples are the important basis of disease diagnosis, health monitoring, and new drug development. The techniques of chip level cell microscopic imaging and analysis lensfree provide new technical means for modern medicine and life science. In view of the issues like high complexity of preparation method, and low image pixel of imaging, and inability to automatically analyze the cell images, super resolution scanning technology will be studied in this project, first of all, based on the combination of CMOS linear array image sensors and microfluidic technology. With the control of microfluidic technology, the cells flow with a constant speed, which is calculated by the double line array structure to compensate the image distortion induced in scan and splice. And then by adjusting the area of pixel photosensitive and frame frequency, and the angle between linear array and flow direction of cells respectively, the two-dimensional scanning resolution will be raised, and the super-resolution images are stitched. Based on these research, this project will design the gray value contrast stretching quantitative nonlinear circuit design cell area, and improve the precision of the cell gray image, and analyze the cell scan images with deep learning algorithms, simplify the architecture of CPU+ acceleration component, implement configurable algorithms of deep learning, and complete the research the cell microscopic imaging and image analysis chip. The study in this project lies a foundation for the realization of chip level microscopic imaging lensfree and image analysis system, and has very broad application prospects in many aspects, such as the wise health, community health care, family health care, and remote medical treatment, and so on.

对生物细胞图像进行采集和分析，是疾病诊断、健康监测、新药研制的重要依据。芯片级无透镜细胞显微成像及分析技术为现代医学和生命科学提供了新的技术手段。针对目前无透镜细胞图像显微系统制备方法复杂、成像像素过低和不具备图像自动分析功能等问题，本项目研究线阵CMOS图像传感器与微流控技术相结合的超分辨率扫描技术，采用双线阵结构计算细胞流速，补偿扫描拼接图像畸变；通过调整像素感光区域面积及帧频、线阵与细胞流动方向的夹角分别提高细胞2维扫描图像的分辨率，得到超分辨率拼接图像；在此基础上，设计细胞区域灰度值对比度拉伸非线性量化电路，提高细胞图像灰度精度，采用深度学习算法对细胞扫描图像进行分析，构建简化CPU+加速部件的架构，实现可配置深度学习算法的硬件加速，完成细胞显微成像及图像分析芯片的研究。本项目研究为实现芯片级无透镜显微成像及图像分析系统奠定基础，在智慧医疗、社区医疗等领域具有极为广阔的应用前景。

针对基于CMOS图像传感器的无透镜成像系统分辨低及片上实时细胞分析难以实现的问题，本项目提出了基于CMOS双线阵图像传感器的片上细胞显微成像及处理芯片的研究目标和方案。. 首先构建了基于倾斜单线阵图像传感器的超分辨率细胞扫描成像方法，提出了一种基于双线阵图像传感器的细胞流速实时可测的超分辨率扫描成像方法。建立了综合考虑光源、衍射光路、图像传感器像素尺寸、帧频、细胞样品流速等系统特性的双线阵超分辨率成像系统的扫描成像及图像重构模型。搭建了双线阵超分辨率细胞扫描图像采集系统，设计了二维聚焦的微流控芯片，完成了微球、红细胞和白细胞的采集实验，并利用扫描图像重构模型对采集的图像数据进行了超分辨率重构。结果表明，采用该扫描成像方式，当倾斜角度在15度时，空间分辨率提升了4倍。. 针对无透镜成像系统对线阵CMOS图像传感器的性能需求，提出了一种TDICMOS图像传感器读出电路。采用4T像素单元结构，通过增加暗像元提高了像素单元的一致性，有效去除背景噪声，进一步使用相关双采样电路降低了固定噪声。基于传统单斜坡模数转换器，通过非线性映射方法拉伸细胞区域灰度值对比度，提高细胞局部图像质量。研究高速列并行模数转换器提高CMOS线阵图像传感器的成像速度，设计了12位逐次逼近加单斜坡两步式列并行模数转换器。在UMC180nm的标准工艺库下完成原理图及版图的设计与仿真与流片，测试结果表明该ADC的ENOB为10.43bit，SNR为65dB，SFDR为74.8dB，THD为-73.46dB。. 研究了无透镜成像系统衍射图像的恢复算法，采用卷积神经网络研究了酵母菌自动分割网络、白细胞融合网络及水藻分类网络。进一步研究了以上网络的加速方法。 分别探讨了基于嵌入式SoC及RISC-V的网络加速方案。为了实现藻类细胞分类算法的硬件映射，对分类模型的权值进行量化，以降低其硬件上的存储开销。最终权值量化网络正确率达到95.1%，精确率达到95.56%，召回率为95.22%。. 通过以上研究，本项目目标已达成，为实现细胞片上检测奠定了理论基础。