Classification and diagnosis of cervical cancer with stacked autoencoder and softmax classification

Abstract

Cervical cancer is one of the most common cancer types in the world, which causes many people to lose their lives. Cancer research is of importance since early diagnosis of cancer facilitates clinical applications for patients. The aim of the study is the automatic diagnosis of cervical cancer. For this purpose, a data set containing 668 samples, 30 attributes and 4 target variables (Schiller, Citology, Biopsy and Hinselmann) from the UCI database was used in the training and testing phases. Softmax classification with stacked autoencoder, one of the deep learning methods, has been used to classify data sets. At first, by applying stacked autoencoder to the raw data set, a reduced dimension dataset is obtained. This data set has been trained for classification by applying the softmax layer. In this phase, 70% (468) of the data set was allocated for training and the remaining 30% (200) for testing. In order to compare the classification performance of the softmax classification with stacked autoencoder, decision tree, kNN, SVM, Feed Forward NN, Rotation Forest models, which are machine learning methods, are used. In the study, proposed models were applied separately to 4 target variables of the dataset, and their classification successes were compared. Finally, softmax classification with stacked autoencoder model, which was applied for the first time in the cervical cancer dataset, performed better than the other machine learning methods with a correct classification rate of 97.8%. Given the interest in machine learning methods in cancer research, new methods of diagnosis are presented in this study in terms of patient diagnostic support systems.