Chromosome Anomaly Detection Based On Convolutional Neural Network

Chromosome,as the carrier of genetic material,will lead to abnormal gene expression when its phenotype is abnormal.Therefore,karyotype analysis of chromosome images has become one of the important topics in cytogenetics research and has important practical value for the diagnosis of human genetic diseases.Because the position of chromosome abnormality is not fixed and difficult to predict,and it is impossible to collect enough negative samples for model training or feature extraction,the abnormal detection of chromosome image has important application value.In this thesis the principle of convolution neural network and classical model structure were studied,aiming at the problem of abnormal chromosome negative samples is difficult to collect,is put forward based on the convolution a semi-supervised learning of neural networks in chromosome anomaly detection algorithm.Feature extraction of chromosome image was carried out by convolutional neural network,and the chromosome image was reduced in dimension and reconstructed by auto-encoder structure.The whole model was trained by generative adversarial network.According to its reconstruction error,the abnormality of the input image was predicted,and the input chromosome image was classified.The experimental results show that this method is effective in detecting chromosomal abnormalities by using real chromosome data samples.The accuracy of the improved model can reach 87.8%and the classification accuracy can reach 84.8%.The main contents of this thesis are as follows:(1)in this thesis,the chromosome image made by G-banding technology is taken as the research object,and the chromosome image is preprocessed,including image denoising,image segmentation,edge detection,central axis extraction,modification and extension,etc.(2)due to the non-rigid nature of chromosomes,chromosomes are prone to bending,so it is necessary to straighten them.In the process of straightening,linear interpolation is introduced in the gray projection of chromosomes to improve the straightening effect.(3)in order to test the model,612 chromosome karyotype analysis charts were made into chromatid images with categorylabel,and the images which could not be used for training were deleted,and26204 chromatid images were obtained.(4)the abnormal detection algorithm based on convolutional neural network is improved.During the training,the label information of normal samples can be combined with the label information of normal samples.During the test,the category labels can be predicted while the abnormal detection is carried out,and the chromosomes can be classified and paired to reduce the labor intensity of professional karyotypes.