Classification of single-cell cervical pap smear images using EfficientNet

Cervical cancer is the second leading cause of death in women in developing nations like India. The automated cervical cancer screening systems play a critical role in diagnosing the ailment at an early stage. They can also assist in routine cervical cancer screening and decision-making processes. The underlying binary classifier for detecting cancerous cells is necessary for building computerized screening systems. Thus, this work aims to develop a robust binary classifier capable of classifying single cervical cells as normal and cancerous. The current work aims to improvise the EfficientNet-B7 with empirical resolution parameters and skilful consolidation of the outcomes of the inner layers using the global pooling layer to improve the binary classification result. The proposed model is trained and evaluated on three independent pap smear datasets. The class activation heatmap visualizes the classification model for better interpretability. The performance of the autonomous classification tasks indicates that the deep learning-based binary classifier is robust in classifying cancerous cells as it achieves accuracy of 94% on Herlev Dataset which is 2% higher than the other state-of-the-art methods and comparable accuracy of nearly 99% on other cervical cell datasets using the proposed methodology.