Using convolutional neural networks for cervical cancer classification with pap smear cytological images

Abstract

Objective: To implement a convolutional network architecture to identify disease patterns in different cervical surgery options using Pap smear data, which refer to the images obtained from the exam, which is used to detect changes in cervical cells that may be indicative of cervical cancer. Methods: This is an experimental study, using the Pap smear cytology image dataset in the public SIPaKMeD database, available online in the Kaggle repository. The proposed transfer learning was developed using pre-trained convolutional neural network (CNN) models, such as InceptionV3, Xception, VGG and a modified VGG approach, to classify Pap smear images according to cell categories: Dyskeratotic, Koilocytotic, Metaplastic, Parabasal and Superficial-Intermediate. Results: The best results in classifying cell categories were for the InceptionV3 and VGG architectures with a modification in the layers of the original architecture, which presented 91% accuracy. The Xception architecture presented 88% accuracy and VGG without modification in the layers presented 86% in the trained model. Conclusion: It is concluded that the findings highlight the model's ability to discern patterns and differentiate cells accurately.