Cervical Cancer Diagnosis Using Very Deep Networks over Different Activation Functions

Abstract

Cancer prevention is mainly achieved by screening the transformation zones. Cervical pre-cancerous stages can be seen in three different types, and all can transform into cancer. Thus, it is crucial to intelligently screen cervical abnormality and have a robust system for detecting whether a cervix is in normal (healthy) or at a pre-cancerous stage. Deep learning showed great potentials when applied to biomedical problems, including medical image analysis, disease prediction, and image segmentation. Hence, in this paper, very deep residual learning based networks are designed in order to perform cervical cancer screening. Moreover, in this work, we highlight the importance of the activation functions on a residual network (ResNet)'s performance. Thus, three residual networks of the same structure are built with different activation functions. The employed models are trained and tested using a dataset of colposcopy cervical images, and the experimental results showed that designed residual networks with leaky and parametric rectified linear unit (Leaky-RELU and PRELU) activation functions performed almost equally in terms of accuracy where they reached accuracies of 90.2 and 100%, respectively. This achieved high accuracy was compared to other related works' results, and it showed an outperformance in screening the pre-cancerous and healthy colposcopy cervical images. Such an earlier and accurate diagnosis may help in preventing cervical cancer transformation.