Study On Segmentation Method Of Overlapping Chromosome Based On Deep Learning

Chromosome segmentation is one of the main steps in karyotype analysis.Accurate segmentation of overlapping chromosomes is beneficial to the diagnosis of diseases according to their structure and number.At present,the diagnosis of chromosome abnormality mainly relies on the automatic analysis system,but when the chromosome overlaps,it needs to be manually divided into separate chromosome,which is highly subjective and costly.Adding computer-aided diagnosis system to chromosome segmentation can reduce the workload of doctors and improve the accuracy of diagnosis.The traditional image segmentation methods are mainly based on the shallow features of the image,such as gray level,color,texture and shape.Similarity is shown in the same areas,and difference is shown in different areas.These shallow features have poor robustness and cannot fully express the essential features of the image.Deep learning can learn deep features from images,improve the generalization ability of the network,and apply the extracted features to the image segmentation task to improve the accuracy of segmentation results.This paper studies and discusses the problem of overlapping chromosomes segmentation,analyzes the principle of image segmentation,and proposes two effective image segmentation algorithms for overlapping chromosomes.The main work contents are as follows:(1)Segmentation method of chromosome based on residual U-Net.Firstly,the structure of U-Net is adopted,and the residual block is referenced on the convolution layer with the same resolution.Then,the convolution layer is replaced by the dilated convolution,and the convolution at the bottom of U-Net is replaced by the dilated convolution at different scales.The features of different resolutions obtained from the contracting path are fused to improve the skip connection.Finally,Tversky loss function is used to reduce the influence of the imbalance of foreground and background on the segmentation results.The results show that the Interaction of Union(Io U)of the residual U-Net reached 96.36%.It is proved that this network can improve the segmentation precision obviously.(2)Segmentation method of chromosome based on deformable U-Net.Firstly,the structure of U-Net is adopted,and the features extracted by the convolution and deformable convolution are fused with the idea of Inception.Then the input images of different scales are fused with the features extracted from the convolution in the contracting path,and the convolution at the bottom of U-Net is replaced with multi-scale dilated convolution.And the skip connection in the network is replaced with the dense block.Finally,this paper adopts the combination of cross entropy loss function and Io U loss function to measure the difference between the predicted image and the ground truth and reflect the segmentation effect of the predicted image.This paper compares it with the existing classical image segmentation networks.The results show that the Io U of the deformable U-Net reaches 96.93%,which can accurately segment chromosome images.In the paper,the segmentation results of the deformable U-Net,residual U-Net and the U-Net are compared on the noise images with different types and degrees of pollution,and it is proved that the segmentation accuracy of the deformable U-Net is the highest.In addition,this paper reduces the time needed for network training by model compression of the network.The paper has 37 charts,5 tables and 84 references.