Research On Automatic Segmentation Arrangement And Abnormal Diagnosis Of Human Chromosome Image

At present,the human chromosome karyotype analysis system used by the People's Hospital of Zhejiang Province has greatly reduced the labor of labor and improved the accuracy of analysis and judgment compared with the traditional artificial karyotype analysis.However,there are also problems such as manual segmentation of human overlapping chromosomes and low accuracy of automatic chromosome classification,and in the diagnosis of abnormalities of human chromosomes,it is still at the level of diagnosis based on professional knowledge and years of accumulated experience of doctors.Make a diagnosis.Therefore,in view of the above problems and the actual situation,this paper studies the problem of automatically solving the automatic segmentation of overlapping chromosomes,improving the accuracy of automatic chromosome classification and automatic diagnosis of human chromosome abnormalities.The main work content and results are as follows:(1)The human chromosome image is initially extracted according to the breadth-first search idea to obtain the chromosome to be identified,including the chromosome monomer and the overlapping and blocking multi-chromosomes.Aiming at the problem of automatic segmentation of overlapping chromosomes,this paper proposes a method to combine the curvature and angle of overlapping chromosome contour points,which can effectively complete the automatic segmentation of overlapping chromosomes.(2)A human chromosome image contains 46 chromosome monomers,and 20700 samples of chromosomes are finally obtained by segmenting and extracting 450 human chromosome images.(3)The classification of human chromosomes is mainly based on geometric features(relative length of chromosomes,centromere position and arm ratio,etc)and banding features.This paper constructs a siamese network structure suitable for human chromosome classification.The siamese network uses the convolutional neural network for feature extraction and uses the loss function of the eigenvector similarity to achieve automatic classification of human chromosomes.In view of the fact that the feature extraction of curved chromosomes affects the classification accuracy,this paper proposes a straighteningalgorithm for curved chromosomes to straighten it.In order to improve the classification accuracy of human chromosomes,the siamese network has been improved as follows: 1)replacing the original single-level convolution operation with a multi-level convolution feature extraction structure;2)replacing the pooled layer before the fully connected layer with the spatial pyramid pooling layer improves the speed and accuracy of human chromosome feature extraction.The experimental comparison results show that the improved siamese network has an accuracy rate of 85.32% for automatic chromosome classification,which is better than other methods.(4)In view of the abnormal judgment of human chromosomes,in view of the advantages of convolutional neural network in feature extraction and One-Class SVM can deal with the imbalance of human normal chromosomes and abnormal chromosome samples,this paper proposes a combination of convolutional neural networks and One-Class SVM human chromosome automatic anomaly detection algorithm.The experimental comparison shows that the applicability of the human chromosome automatic anomaly detection method based on the karyotype analysis technology can solve the problem of current system visual diagnosis,and the accuracy rate is up to 89.5%.