Super-resolution Processing And Classification And Recognition Of Cell Imaging Based On Deep Learning

Cells are the most basic structural unit of the human body.Cell detection,such as circulating tumor cell detection and complete blood count,are medical detection methods that are widely used to assess cancer and immune conditions in patients.The current cell detection methods include manual microscope detection and cell detector detection represented by flow cytometry,but these two methods have some shortcomings: manual microscope detection is time-consuming,inefficient,and requires an experienced doctor to judge;Automated cell detectors are often costly and are not suitable for remote areas or areas with underdeveloped medical resources.Therefore,it is necessary to develop a miniaturized,low-cost,and intelligent microfluidic cell detection device.However,the system structure of the miniaturized microfluidic cell imaging detection device is simple,which makes it difficult to collect high-resolution cell images.So,further intelligent analysis based on cell images requires efficient recognition and classification processing algorithms.Based on the deep learning method,this paper proposes and implements the super-resolution processing of low-pixel cell images based on parallel residual super-resolution network(PRSRCN)and the three-class recognition of label-free(unstained)white blood cell images based on Resnet-50.(1)The parallel residual super-resolution network PRSRCN uses serial residual blocks and parallel convolutional layers for feature value extraction,and sub-pixel convolutional layers for high-resolution image reconstruction.The jump connection mode of the residual blocks used in the PRSRCN network can increase the model's adaptability to low-pixel cell images,integrate the original cell information into the newly extracted cell information,and parallel convolution operations can expand the width of the network.Compared with super-resolution networks such as SRCNN and VDSR,the jump connection mode and parallel convolution operation of the PRSRCN network can better extract low-pixel cell information,and the PSNR(Peak signal-tonoise ratio)index of the PRSRCN network is higher than other networks by more than0.3,and the SSIM(Structual similarity)index is higher.Other networks are above 0.01,which proves that the PRSRCN network has a good recovery effect on low-pixel cell images.(2)Aiming at the characteristics of the cell images collected by the miniaturized label-free microfluidic cell imaging detection device,a Resnet-50 network based on migration learning is proposed to classify label-free white blood cells into three categories.Transfer learning can quickly optimize network parameters.Compared with the Inception V3 model without short connection structure,the short connection structure of Resnet-50 is more suitable for feature extraction of unlabeled white blood cells,and the Resnet-50 network can reach 94%Accuracy in the testing set.The labelfree white blood cell dataset is collected through a microscope and directly trained through the network,which reduces the white blood cell staining process and avoids the effect of the dye on the physiological state of the white blood cell.The super-resolution method for low-pixel cells and the three-classification method for label-free white blood cells proposed in this paper provide a potential solution for the realization of miniaturized cell detection instruments.