An Improved Transfer Learning-Based Approach for Detecting Abnormal Cervical Cells from Pap Smear Microscopic Images

Abstract

Cervical cancer is one of the most common and dangerous diseases for women's health. Automated abnormal cell screening will help detect early and improve the accuracy of cervical cancer diagnosis, from which women can plan treatment, preventing the development of cervical cancer and improving survival rates. This paper proposes an improved method to detect abnormal cervical cells from images of cells that are stained and examined under a microscope based on a transfer learning approach. Achieved results show that the proposed EfficientNetv2L-SVM model has an accuracy of 92% and an F1-score of 88%, higher than the ResNet50v2 deep learning model with 90% and 83%, respectively. The measurement of Precision and Recall also gives the same result. Experimental data were collected from the Cytopathology Laboratory of Binh Dinh province over ten years, from 2014 to 2023 with five different types of cervical cells based on the diagnosis of expert pathologists (The dataset in this research is provided and allowed to use by the Cytopathology Laboratory, located at Quy Nhon City, Binh Dinh Province, VietNam.).