Cervical Cell Image Classification Based On Multiple Attention Fusion

Abstract

Accurate classification of cervical cells is of great significance to the detection and treatment of cervical cancer. Over the years, convolutional neural network (CNN) has been successfully applied to cervical cell classification. Recently, the attention mechanism has been the research focus, which can learn local discriminant features. To further improve the performance of cervical cell image classification, we propose a novel cervical cell image classification method based on multiple attention fusion in this paper. Specifically, the Squeeze and Excitation (SE) and Spatial Attention Module (SAM) blocks are fused to learn the dependency between features from the channel and spatial directions respectively. In order to capture the long-range dependencies between features, the features embedded with SE and SAM are further fused with Disentangled Non-Local block (DNL). Experimental results on the publicly available cervical cell dataset SIPaKMeD show the effectiveness of our method.