Classification Of Cervical Intraepithelial Neoplasia Based On Colposcopic Images

In worldwide,cervical cancer ranks as the fourth leading cause of cancer death in women,more severe in some resource-poor areas,cervical cancer morbidity and mortality ranks second,preceded only by breast cancer.However,as a main screening tool,visual inspection with acetic acid relies remains dependent on operators’ subjective interpretation of the acquired images and still remains lower diagnostic accuracy.Therefore,how to achieve automated diagnosis and improve screening efficiency and accuracy in this screening process,is the main research direction of most researchers.Despite the increasing number of automated classification studies associated with cervical intraepithelial neoplasia,most of the current work aimed to distinguish patients with high-grade lesions from those with low-grade lesions and normal ones,but few studies on classification of non-pathological(normal)and pathological(including high-grade lesions and low-grade lesions)are involved.In that context,this paper proposed an automatic classification algorithm based on lesions and non-lesions.The study mainly included the color or texture features commonly used in the classification of cervical intraepithelial neoplasia and four commonly used classifiers(support vector machine,random Forest,back propagation neural network and Knearest neighbor).For each classifier,the specific optimal combination was selected by means of feature selection,and conducted a unified assessment of the performance of the four classification models(classifier + best feature combination).Firstly,the cervical region was extracted by k-means clustering algorithm,and then the possible reflection points in the cervical region were removed,and the cervical images before and after the acetic acid experiment were registered by cross-correlation image matching method,and the cervical region overlaps of the two images were obtained.Secondly,the evaluation of the whitening level was completed in the overlapping cervical region.On this basis,two color features and five texture features based on the gray co-occurrence matrix were extracted to form the feature set,and the data set used in this paper was constructed by combining diagnostic opinions of the expert.Finally,for the seven colors or texture features mentioned above,in combination with he feature selection method of exhaustive search and cross-validation to obtain the specific optimal feature subset of the four classifiers.The classification performance of the four classifiers under the specific optimal feature combination was tested in the same test set.The proposed method was validated on 350 patient data sets(including 175 positive patient and 175 negative patient).The results shown that the support vector machine achieves 81.4% accuracy,82.8% sensitivity and 80.0% specificity under the specific optimal feature subset,outperforms the other three classifiers.The finding of this study may provide useful reference values to the development of an automatic cervical cancer screening tool.