Metaheuristic Driven Framework for Classifying Cervical Cancer on Smear Images Using Deep Learning Approach

Abstract

Cervical cancer, among the most frequent malignancies in women, frequently manifests after a protracted and curable precursory phase. Its effects and mortality are especially severe in economically weak regions. This emphasizes the critical need for rapid investigation and creation of methods to improve detection rates, which in turn improve treatment effectiveness, patient survival rates, and reduce burden of healthcare. The difficulties of extremely fine vision classifying are intrinsically linked to the complexities of cervical cancer diagnosis. Conventional techniques for classification used to diagnose cervical carcinoma frequently rely on cellular segmentation and methods to extract features. In contrast, Convolutional neural network (CNN) models need a substantial dataset to address issues of over-fitting and inadequate generalization. In order to attain better classification accuracy and simplified processing temporalities, this work intends to create modified deep learning models for automated cervical cancer diagnosis that do not rely on segmentation techniques or customized features. Due to lack of training data, this work used transfer learning with pre trained models that had already been trained to perform binary class classification on pap smear images. By utilizing the open-source Herlev’s dataset, this work conducted a comprehensive analysis and comparison of four pre-trained deep CNN models. Notably, the accuracy and computational time achieved by proposed technique is 99.65% with better specificity as well sensitivity and 169 seconds respectively. Two independent meta-heuristics have confirmed the framework’s usefulness in supporting cervical cancer diagnostic testing professionals.

Keywords

• Cervical cancer
• feature selection
• meta-heuristics
• contrast enhancement
• support vector machine