Research On Cell Segmentation Algorithm Based On Deep Learning

Fluorescent in situ hybridization(FISH)is a molecular cytogenetic technique that provides reliable imaging biomarkers to diagnose cancer and genetic disorders in the cellular level.One prerequisite step to identify carcinoma cells in FISH images is to accurately segment cells,so as to quantify DNA/RNA signals within each cell.However,for FISH cell images,the artificial annotation data is extremely valuable,so only a small amount of annotation data can be obtained.In addition,during the automatic segmentation of FISH cell images,it is usually difficult for many automatic segmentation algorithms to accurately segment target cells due to the uneven staining of staining agents,low image contrast,weak cell boundaries,and cell touching.In view of the above problems,we conduct research on FISH cell image segmentation,and the main work is as follows:1)A deep convolution segmentation algorithm for FISH cell images is proposed.Firstly,in order to improve the segmentation speed of the convolutional neural network model,1 × 1 convolution kernel and the appropriate number of convolutional channels are cleverly used to reduce the model parameters,thus improving the efficiency of cell segmentation.Secondly,in order to accurately segment the FISH cell images,we incorporate the underlying feature information in the original images into the symmetric network model and improve the relevant loss function,thus improving the segmentation accuracy of the cell images.Finally,the results of image segmentation are further improved by improving the loss function.2)A watershed algorithm based on the center point of the cell is proposed as a postprocessing,which can effectively separate the areas of cells that adhere to each other.In the results of cell segmentation,the target cells may adhere to each other.Different from the U-Net,which uses weight preprocessing to guide the loss function to separate the cell boundaries,we first train the network model of cell center with deep learning to identify the cell center.Subsequently,we use the identified cell center point as a priori guidance information and combine the watershed algorithm to isolate the adhering cells.With this kind of post-processing,the adhesive cells can be separated effectively.3)A semi-supervised cell segmentation algorithm based on Generative Adversarial Nets(GANs)is proposed.Labeling medical cell images is an extremely tedious process,so we use a semi-supervised learning strategy to reduce the reliance of convolutional neural networks on large amounts of labeled data.In the training process,using the Generative Adversarial Nets forces the generator(segmentation network)and discriminator(discrimination network)to learn against each other,so that the segmentation network can use more unlabeled cell data to improve segmentation accuracy.