Two-stage classification of pap-smear images based on deep learning

After years of discovery, the cancer cervix is still a significant worldwide threat that can be detected early using the pap-smear test. The pap-smear is a screening procedure to find a candidate or positive cancer cell. Recently this process has been done using deep learning, especially Convolution Neural Networks (CNN). The Herlev dataset with seven class data is one of the public datasets that has been researched. Since the high similarity of pap-smear images, previous research has modified the data into two large categories to provide a good result. However, they still require some improvement in the original seven-class classification. We proposed the two-stage classification based on deep learning on pap-smear images to specifically classified the data into their original categories. This method classified the dataset into five classes, then reclassified them into three. In the end, the dataset has been classified into seven classes following the original dataset. This research uses various types of CNN, such as VGG types, ResNet, MobileNet, and EfficientNet. As a result, this proposed method gives 78.73% accuracy in test data. This result increased by 26.77%, better than using the one-stage classification. Due to the data augmentation technique, this method provides an accuracy of 90.98% by combining three data types. In further research, this method could be an idea to classify other high-similarity data cases, such as medical images.