Deep Learning for Cervical Cancer Detection: Enhancing Accuracy with Adaptive Test-Time Augmentation (ATTA)

Cervical cancer detection in the images of the Papanicolaou (Pap) smear test is an essential way to find out the primary stages of the disease. It is sad to admit that deep learning models face the challenge of generalization of the results due to the lack of data. This research suggests a new model based on Xception to enhance the accuracy of cervical cancer classification. Using SIPaKMeD, a cervical cytology dataset application with data augmentation, batch normalization, and random oversampling (ROS) are expected to help improve the performance. The test accuracy of the proposed model will be 95%, which is three percent higher than the traditional deep learning model. The results will show that ATTA helps to increase model robustness and have fewer misclassification errors.