Automatic Classification Of Human Chromosome Images

Chromosome image classification is one of the key steps in clinical chromosome analysis,which is of great significance in the diagnosis and oncology of genetic diseases.Due to the development and progress of computer technology,computer-based chromosome classification has become a hot research in recent years.Chromosomes are non-rigid objects,which are prone to bending,and bent chromosomes will affect the accuracy of the network,and need to be straightened for bent chromosomes.The existing chromosome straightening methods are mainly divided into cut straightening method and skeletal association straightening method,which have some limitations:cut straightening method,which straightens chromosomes by cutting pictures,will make the image information at the cut position missing and destroy the coherence of chromosomes;skeletal association straightening method,which straightens chromosomes by associating chromosome pixels with bones.However,it is susceptible to intermediate steps,which will change the chromosome morphological information if the bones are not extracted accurately or the chromosome pixels are not associated with the bones accurately.In addition,chromosome data has certain particularity.The difference of chromosome morphology between adjacent classes is small,while the difference of chromosome morphology between far away classes is large,that is,the similarity of chromosome morphology between adjacent classes is higher,which leads to the classification confusion of samples adjacent to classes.In addition,the banding characteristics of the chromosome at the bottom of the category are fuzzy,and the classification accuracy is lower than that of the chromosome at the top of the category.For the above problems,this paper puts forward targeted improvement,the main contributions are as follows1.Aiming at the problem of chromosome bending,a new method of chromosome straightening is proposed,which uses circle fitting to straighten the bent chromosome.The straightening method proposed in this paper avoids the image missing problem of cutting straightening method.At the same time,it avoids the chromosome morphological information change problem of bone association straightening method.2.In order to concentrate the advantages of different straightening methods,a multi input network feature extraction network integrating multiple straightening methods is proposed.A variety of straightening images are input to improve the classification result of the network.3.Due to the higher similarity of chromosome morphology between adjacent classes,an improved loss function based on softmax loss is proposed,which is called "repul-center loss".It measure the distance between the sample and the feature center,and punish when the sample feature is closer to the feature center of adjacent classes.4.In order to solve the problem that the information of chromosome banding at the bottom of the classification is fuzzy,a prediction branch of the network is added to predict the grouping of chromosomes,assist the existing chromosome classification branches,and enhance the accuracy of classification.Finally,combined with all the methods proposed in this paper,we get better classification results than other existing methods on the BioImLab public dataset.