Layer Selection on Residual Network for Feature Extraction of Pap Smear Images

Abstract

Pap smear screening test is one of the early prevention efforts to detect cervical cancer. Manual screening tests are still prone to observation errors. This study aims to create a convolutional neural network (CNN) model and support vector machine (SVM) model to identify cervical cancer through pap smear images. The data used are 4049 normal and pathological cervical cells in pap smear images sourced from SIPaKMeD, which were divided into 5 classes based on the level of cancer malignancy. The CNN model is used to extract features on the pap smear image, and SVM is used to carry out the classification. The results of this study are four cervical cancer classification models on pap smear images using Resnet50 and Resnet50V2 architecture and SVM algorithms with different scenarios on freeze and unfreeze of the convolution layer. The classification model with the best performance has an accuracy of 97.09%. CNN model with freezing the convolution layer provides much faster in the pre-trained model and the integration of this model with the SVM as the classifier results in the classification model of cervical cells in pap smear images with high accuracy.