Synthesis And Classification Of Overlapping Cell Images Under Imbalance Data

In recent years,with the rise incidence of cervical cancer worldwide,cervical cancer has become a social problem which threatens women's lives.Recently,DNA ploidy analysis of cells is an automation technique applied in pathological diagnosis.It effectively combine the diagnostic experience of pathologists with the precise calculation and rapid processing capabilities to identify cervical cell images.Firstly human cell specimens are collected and the DNA in cells is stained,then the cells are placed under a microscope to obtain nuclear images through a high-resolution camera.The model training method is used to classify the nuclear images and identify various types of nuclei,then image processing techniques are used to measure the relative content of cellular DNA.Finally,the abnormal cells are listed to assist the doctor in diagnosis.However,Abnormal cells mostly exist in overlapping cell clusters,so it is important for the DNA ploidy system to accurately identify overlapping nuclei.Due to the wide variety of overlapping nuclei,it takes time and effort to collect a large number of overlapping cell nuclei,which makes the number of overlapping cell images much smaller than that of other images,resulting in imbalance training data.However,most classifier learning algorithms are not suit for imbalance data.Therefore,it is of great significance solving the problem of imbalance training data for the DNA ploidy system.In order to solve the above problems,synthesis and classification of overlapping cell images un der imbalance data was proposed.This paper includes the following aspects:Firstly,An original cell image selection method based on cosine invariance is proposed.The method first extracts the feature values of the cell image,and then selects representative single cell images for synthesizing a large number of overlapping cell images based on the feature space using the principle of cosine invariance.When the angle formed between the samples is less than the threshold,only one input sample image is used for the overlapping cell images,increasing the diversity of the synthetic data samples.Secondly,a randomness based method for synthesizing overlapping cell images is proposed.Firstly,new overlapping cell images are produced by rotation and segmentation,and overlapping cell images are synthesized by two single cell images.In order to make the synthesized cells as close to real as possible,three main problems are considered.To ensure that the synthesized cells are representative,single cells are selected to obtain a typical single cell image for synthesis.In order to avoid excessive aggregation of the synthesized data in sample space,the randomness is introduced in both the angle of rotation and the overlapping degree of the cells.In order to make the overlapping part true,the overlapping part of the pixels needs to be reconstructed.Finally,An active learning-based image selection method for post-synthesis overlapping cell clusters was proposed.The method first divides a large number of samples into multiple sample clusters,increasing the efficiency of actively selecting samples.Secondly,the training samples are selected according to two selection criteria,namely representativeness and uncertainty;then each sample cluster is in the loop iteration process.Neural network model training is performed to select representative training samples;all the overlapping cell images and single cell images will be selected for model training.Experiments show that the recognition rate is improved by adding new cells to the small categories.The method can be applied to the imbalance data problem.When the amount of small sample data is too small to make the classifier learn enough,the images needed for synthesis can simulate the texture,gray scale,shape,and size,thus making the classifier more fully and comprehensively to learn the features of small samples.Finally,combined with the active learning method,high-quality overlapping cell images are selected,and the model learning is better.