Cervical Cancer Segmentation and Classification from Pap Smear Images using Deep Neural Networks

Abstract

Cervical cancer is a significant worldwide cause of mortality, despite its preventability and treatability with early removal of affected tissues. Cervical screening programs must be universally accessible and efficiently implemented, a challenging endeavor that requires, among other factors, the identification of the most susceptible segments of the population. This paper introduces an efficient deep-learning approach for the classification of multi-class cervical cancer using Pap smear pictures. The enhanced SE-ResNet152 model, based on transfer learning, is used for efficient multi-class categorization of Pap smear images.  
The suggested network model accurately extracts trustworthy and relevant picture characteristics. The network's hyperparameters are tuned with the Deer Hunting Optimization (DHO) technique. Eleven classifications for cervical cancer illnesses include five categories from the SIPaKMeD dataset and six categories from the CRIC dataset. A dataset of Pap smear pictures including 8,838 images with diverse class distributions is used to assess the suggested methodology. The use of the cost-sensitive loss function during the classifier's training addresses the dataset's imbalance. The suggested technique achieves 99.68% accuracy, 98.82% precision, 97.86% recall, and 98.64% F1-Score on the test set, surpassing previous methods in multi-class Pap smear picture classification. The suggested technique yields superior identification outcomes for the automated preliminary diagnosis of cervical cancer in hospitals and clinics, attributable to its favourable categorization findings.