Effective cervical intraepithelial neoplasia classification through polarization-based machine learning

Identification of cervical intraepithelial neoplasia (CIN) tissues holds significant clinical importance in reducing the incidence and mortality of cervical cancer. The ultraviolet Mueller matrix imaging polarimeter (UV-MMIP) can significantly enhance morphological specific contrast. For the first time, to our best knowledge, we used polarimetric metrics measured by UV-MMIP combined with machine-learning techniques to achieve high-accuracy CIN classification. Initially, eight classifiers of four types were trained, and the best-performing one was identified, achieving a classification accuracy (F1 score) of up to 0.815. Subsequently, the best classifier from each type was assembled to construct a stacked model to further explore the potential of the machine-learning techniques, resulting in an improved F1 score of 0.838. Additionally, the classification results indicate that depolarization-related metrics have stronger discriminatory power on the trained classifier’s outcome than retardance-related metrics, which highlights the contribution of UV-MMIP for the classification task. Our work validates the feasibility of the polarization-based machine-learning framework for effective CIN classification.