Research On Low Resolution Cell Images Segmentation Based On Convolutional Neural Networks

An on-chip lensless cell detection system based on microfluidic chip and CMOS image sensor provides a new direction for portable and intelligent cell detection equipment.The lens image acquired by this lensless system has low resolution,and the image gray scale change is greatly affected by the experimental environment.The traditional method for cell segmentation is sensitive to noise and not robust,and it is difficult to achieve high-precision segmentation.Therefore,the high-precision segmentation method for studying low resolution cell images is of great significance for segmenting the cell images acquired by the lensless cell detection system.For the automatic segmentation of low-resolution cell images,a method using convolutional neural networks is proposed.Since the cell image acquired by the lensless system is difficult to produce the data set required by the network,the cell image acquired by the simulation system is used to make the data set.The use of low resolution image data sets for network structure design is inefficient.Because the features of high-resolution cell images are obvious,they are easy to segment.Therefore,high-resolution image data sets are used to determine the appropriate initial network structure to shorten the design cycle of the neural network.The initial network model is first adjusted through a high-resolution data set to achieve lightweight networking for later migration.After testing,the segmentation accuracy of the lightweighted network for high-resolution cell images can reach about 97%;Then,using the low resolution data set to verify the feasibility of the above network,the test segmentation accuracy is about 91%,which indicates that the network model determined by the high resolution data set is suitable for low resolution images;Then,because the image of the cell acquired by the lensless system has a diffraction pattern,the network is improved.By analyzing the results of different network structures on the segmentation of the low-resolution diffraction image,the residual structure and the cavity are added to the network structure which can improve the accuracy of convolutional neural segmentation;Finally,the saved network data is analyzed and quantified,the network parameters are reduced,and the network performance is not reduced,which provides a theoretical basis for the realization of portable cell detection equipment.Based on the keras environment,this study runs and tests the proposed neural network cell image segmentation model,and the segmentation accuracy rate is about 96%.And the network data is compressed,the compressed network parameters are reduced by more than 47%,and the segmentation accuracy is about 92%,achieving the expected target requirements.