MEFA-Net: A multi-scale edge features aggregate network for cervical cell semantic segmentation

Leixin Chen, Zhiyong Li, Jin Tang

Abstract: Cervical cancer is one of the female diseases with the highest morbidity and mortality, Early screening of cervical cancer is the key to reducing the incidence of cervical cancer and timely treatment. Several studies have proposed a series of methods for cervical cell semantic segmentation. However, they ignored the importance of edge information. In this paper, a multi-scale edge features aggregate network for cervical cell semantic segmentation is proposed (MEFA-Net). Specifically, we use Canny operator and U-Net network to extract edge features, and use them as prior information in the encoding and decoding stages of MEFA-Net respectively. MEFA-Net uses U-Net as its basic architecture. Multi-scale feature extraction module (MFE) is designed as encoder to improve the multi-scale feature extraction ability of the network and make the network pay more attention to the spatial gradient change area. Secondly, edge feature aggregation module (EFA) is designed to aggregate semantic features and edge features on multiple scales.Experiments on the common data set CCEDD show that MEFA-Net achieves the most advanced performance.