Cervical Cancer Classification Using Deep Learning Approach Using Colposcopy Images
Sundaranarayana Dayalane, Sankar Murugesan, Sandeep Kumar Mathivanan, Hariharan Rajadurai, Karthikeyan Panneer Selvam, Mohd Asif Shah
Abstract
The conventional approach of cervical cancer classification mostly depends on pathologists' expertise, which is typically less precise. Over the last 75 years, the incidence and death rates of cervical cance6r have decreased dramatically thanks in large part to the widespread use of colposcopy, an essential component of cervical cancer prevention. On the other hand, misdiagnoses and a decline in diagnostic efficacy have resulted from the increasing workload in visual screening. In deep learning, cervical cancer type classification has shown improved performance using medical image processing and Convolutional Neural Network (ConvNet) models, namely the ResNet50 model and Cervix Ensemble Network (CervixNET). In order to automatically identify cervical cancer from colposcopy images, this research presents two ConvNet architectures. In one architecture, ResNet50 functions as a transfer learning (TL) model, and for classification, a new model called CervixNET is created. Both models' sensitivity, specificity, and accuracy are evaluated. With an accuracy of 82.67%, ResNet50 produces findings that are quite excellent. For ResNet50, the kappa value suggests a moderate categorization. CervixNET, however, performs very well; its sensitivity, specificity, and kappa score are 99.58%, 99.63%, and 99.12%, respectively, in the experimental data. Notably, the CervixNET model's classification accuracy of 99.23% marks a notable increase over the ResNet50 (TL) model by 16.56%.