中国科学院自动化研究所机构知识库

        细胞计数在基础科学研究和临床医学领域均具有重要的意义。在基础科学研究中，介观图谱构建需要对注射逆行示踪剂的大脑的各个区域进行细胞计数， 通过细胞的数量反映各区域的连接强度。而在临床医学领域，一些常规的检查需要对样本特定的某种细胞进行计数，例如血常规中的红细胞与白细胞数目等。 而在当前的图谱绘制领域，细胞数量主要通过人工计数的方法获得，大量的计数人员对神经细胞图像进行计数，不但需要大量的时间和人力物力，同时多人计数情况下，个体之间的标记误差不可避免。这些问题直接或间接的影响了介观脑图谱的绘制，因此开发基于荧光显微图像的自动化细胞计数方法在提升细胞识别的准确率、提高细胞计数的速度、降低人工成本等方面具有重要的意义。 

       本论文提出了从数据获取到数据处理最后进行细胞计数的一系列流程和方法。数据获取阶段提出了ROI选取算法通过选取单层图像的ROI进而对ROI区域进行多层扫描来提升扫描的速度并减小空间的占用。同时该算法提出的亮度与颜色自适应调节方法同样应用于之后深度学习的训练数据的获取，保障了图像的清晰。在细胞计数阶段，本文设计了基于注意力机制和对抗网络的深度学习结构，在经典的 U-Net 上引入结合空洞卷积的特征增强模块、基于通道注意力机制的特征选择模块、基于注意力机制的像素级特征融合模块，设计了本文的密度图回归网络 DCAU-Net，最后使用对抗网络进行约束保证密度图分布的一致性得到训练的整体框架 ADCAU-Net。为了验证我们的细胞计数自动化算法的有效性， 我们在合成细菌细胞、骨髓细胞、结肠癌细胞和猕猴脑神经细胞四种细胞的数据集上进行了验证并与 U-Net、FCRN、Count-Ception 三种算法进行了对比，结果表明我们的算法在大多数数据集中表现均好于其他算法，尤其是猕猴脑神经细胞图像。最后我们对论文中提出的各模块进行了消融实验说明我们的模型各模块的重要性和必要性。 

       我们提出的算法可以准确、快速、有效的自动化预测细胞数量，为各种需要细胞计数的科学研究提供方法支持，为科研人员带来了极大便利。

     Cell counting plays an important role in basic scientific research and clinical medicine. In basic scientific research, the construction of mesoscopic atlas needs to count the cells in each area of the brain injected with retrograde tracer,and reflect the connection strength of each region by the number of cells. In the field of clinical medicine, some routine blood tests need to count certain cells in the sample, such as the number of red blood cells and white blood cells. In the current field of atlas illustrating, the number of cells is mainly obtained by manual counting. It takes a lot of time, manpower and material resources because a large number of counting personnel to count the fluorescent cell images. At the same time, in the case of multi person counting, the labeling error between individuals is inevitable. These problems directly or indirectly affect the illustrating of mesoscopic brain atlas. Therefore, the development of automatic cell counting method based on fluorescence microscopic image is of great significance in improving the accuracy of cell recognition, accelerating cell counting, and reducing the labor cost. 

    In this paper, we present a series of processes and methods from data acquisition to cell counting. In the data acquisition stage, the ROI selection algorithm is proposed, which can improve the scanning speed and reduce the space occupation by selecting the ROI of single-layer images and then scanning the ROI area to get the multi-layer images. At the same time, the adaptive adjustment method of brightness and color proposed,the algorithm is also applied to get the training data for deep learning, which ensures get the clear image. In this paper, we design a deep learning structure based on attention mechanism and adverserial network. In this architecture, we design density map regression network: DCAU-Net,which combines feature enhancement module based on dilated convolution, feature selection module and pixel-level feature fusion module, both of them based on attention mechanism. Finally, the whole training framework ADCAU-Net is obtained by using the adverserial network to guarantee the consistency of density distribution. In order to verify the effectiveness of our automatic cell counting algorithm, we test it on the 4 datasets which of them is synthetic bacterial cells, bone marrow cells, colorectal cancer cells and macaque monkey brain fluorescent cells scaned by ScanZ1, and compare it with U-Net, FCRN and Count-Ception algorithms. The results show that our algorithm performs better than other algorithms in most datasets, especially macaque monkey brain fluorescent cell image. Finally, the ablation experiments are carried out to illustrate the importance and necessity of each module of our model. 

    Our algorithm can accurately, quickly and effectively automatically predict the number of cells, and provide support for various scientific research which need cell count. It brings great convenience for researchers.