Graph-based Annotation Reliability Estimation in Cervical Cytology

We formulate the transformed label propagation that is recently proposed as a graph-based reliability estimation methodology for existing human annotation errors in labels of cervical cytology image datasets, and we analyze its performance with precision, recall, and label entropy measures. The human annotation error causes the deterioration of a trained model of machine learning model, called label noise. It is critical to the subsequent training phase of the multiple-image classification, which is one of the imminent issues in the medical field. In this work, we drill down into the algorithm of the transformed label propagation, explain how it works, and confirm performance features for label noise detection in both single-cell and cluster-cell images from Herlev and SIPaKMeD Pap smear cervical cytology datasets. Our findings in this study suggested that the transformed label propagation is a powerful tool to detect label noise with quite high accuracy and effectiveness in lowering the label noise of the dataset.