A Robust Deep Learning and Feature Fusion-based Multi-class Classification of Cervical Cells

Abstract

Cervical cancer is one of the prevalent and lethal diseases in women which can be prevented if routine screenings are conducted to find premalignant lesions at an initial stage to cure them. The pap-smear test is used for early detection but this is not much efficient because it is a time-consuming process and other manual screening methods give high human error rates and false-positive results. To overcome the classification of cervical cell problems, a deep learning-based automatic computer-aided diagnostic system has been developed using pap-smear images of cervical cells. This study is focused on different limitations in the classification of cervical cells and a deep learning-based feature fusion network is proposed in this work. The proposed network model has been tested on a publicly available CRIC searchable image dataset. The trained model with an accuracy of 96.07%, 93.30%, and 85.07% on an unseen test set images for 2-class, 3-class, and 6-class classification, respectively. The performance metrics supported the models' accuracy with improved outcomes as compared to the trained single network. The model includes features extracted from various networks, highly efficient feature extraction suitable for cervical cell image analysis and other biological applications.