Feature Pyramid Based Attention for Cervical Image Classification

Abstract

It is known that cervical cancer has been a great threaten to the health of women worldwide. Regular examination through images obtained by colposcopes can facilitate early detection and treatment. However, it is challenging for computer-aided methods to perform diagnosis with the cervical images correctly. To this end, we develop a novel deep neural network based method for cervical image classification in this paper. The proposed method first generates a sequence of feature maps through lateral connection and feature fusion so as to fully exploit the multi-scale information included in a main network. Then compatibility scores are computed between the multiple feature maps and the global feature outputted by the main network, with which attention maps can be obtained. The attention maps are able to extract salient features from images via focusing on the crucial regions of the cervixes, which can be a guidance for clinical diagnosis. We test the proposed method on CIFAR datasets and a cervical dataset collected from the Peking University First Hospital and results show that the proposed method achieves excellent performances and outperforms the related approaches.