Cervical Cell Classification Using Multi-Scale Feature Fusion And Channel-Wise Cross-Attention

Cervical cancer is one of the prevalent malignant tumors in women, and accurate cervical cell classification is clinically significant for early screening of cervical cancer. In this paper, we propose a novel cervical cell classification method based on multi-scale feature fusion and channel-wise cross-attention. Specifically, the multi-scale cell features are combined from the perspective of channels, and then the fused multi-scale features are fed into multi-head channelwise cross-attention to explore the channel dependencies and non-local semantic information, which are encoded into the high-level CNN features through Multi-Layer Perceptron (MLP) with residual structure. More importantly, the Re-Attention is applied to exploit the correlation among different attention heads. Experiments on three public cervical cell datasets, SIPaKMeD, Herlev and Motic, demonstrate the effectiveness of the method for cervical cell classification.