Research On Cervical Cell Classification Technique Based On Machine Learning

Cervical cancer is the second most common cancer among women,with more than 250000 deaths every year,especially in developing countries,there are a large number of new cases every year,which is about 1/4 from the China.Although the incidence of cervical cancer is high,but fortunately,cervical cancer has a relatively long period in the early stage of the disease,which makes early screening and timely treatment can increase the cure rate,so the cervical cancer screening is very important for the health of women.There are many cervical cancer screening methods,one of the most commonly used method is Pap smear detection,however,this method relies on manual reading,which will cause a series of problems,such as the physician workload,the high cost of screening,screening results of the reliability and accuracy is influenced by the work of physicians professional technology and subjective emotions etc..Therefore,the study of automated system combined computer technology with cervical cancer screening for the prevention and treatment of cervical cancer is of great significance.The automation technology of cervical cell image is divided into five steps,which are image preprocessing,segmentation,feature extraction,feature selection and classification.In the automation technology of cervical cell image,many researches focus on image segmentation and feature extraction,and get a good result.However,there are still some shortcomings in the research of feature selection and classification.Therefore,this paper mainly focuses on the feature selection and classification of cervical cell image.The first research focus of this paper is to propose a method of automatic classification of cervical cells based on ReliefF feature selection and random forest classification.In the feature selection stage,the ReliefF method is used to weight the 20 dimension of the Herlev data set;In the classification stage,the random forest method divides the Herlev data into two categories: normal cells and abnormal cells.After cross validation,the experimental results show that the random forest method has the best performance at the top 13 dimension,which is better than Na?ve Bayesian method,C4.5 method and logical regression method.Most of the previous studies are based on the assumption that the training data and test data are identically distributed,however,in the actual application,the distribution of cervical cells may be affected by regional,age and other factors,which makes the original classifier cannot adapt to the new test data,resulting in the classification results have some deviations,so the second research focus of this paper is to apply transfer learning to the classification of cervical cells.The training method can adjust the classifier according to the actual distribution of cervical cells,so that the classification results are more practical.In this paper,two kinds of transfer learning methods in the classification of cervical cells: The first method is cervical cell classification method based on bias correction,according to the distribution of the source domain and the target domain,choosing a data more close to the target domain from the source domain,and then adding it to the target domain as new training data set,and using the new data set training classification model to make it more suitable for the target domain.The second method is cervical cell classification method based on weighted local model,using the different distribution of each sample data in the target domain set different weights for different base classifiers,and then to form a strong classifier.The proposed two methods are tested by cross validation and the experimental results show that these two methods can achieve good classification accuracy in the classification and recognition of cervical cells.