Cervical Lesion Classification via Positive-Unlabeled Learning

Abstract

Cervical lesion classification has gained the attention of researchers because of its importance in the mitigation and treatment of cervical cancer. Most machine learning approaches have addressed the task under the single and fully labeled assumption; however, such assumptions do not reflect the nature of cervical lesions in a clinical setting. In this study, we adapt a semi-supervised learning algorithm on a partially labeled multi-label cervigram dataset by treating it as positive-unlabeled. We simultaneously trained a classifier, and a propensity model to simulate the clinical bias in labeling lesions. Results were compared to that of a supervised learning model, showing improvements on several performance metrics. Gradient-weighted class activation mapping also showed better learning focus.