Binary tree-like network with two-path Fusion Attention Feature for cervical cell nucleus segmentation

Abstract

Early detection of cervical lesion is of great significance in reducing mortality from cervical cancer, and segmentation of cervical cell nuclei plays an important role in screening for cervical lesion. Compared with traditional algorithms, several deep learning methods can improve the segmentation; however, due to blurred boundaries and complex gradients of medical images, the segmentation remains unsatisfactory. In addition, the existing datasets used by cervical cell nucleus segmentation research are lacking in terms of both quantity and diversity, so methods based on those datasets cannot effectively cope with challenging cases. This paper releases a new cervical cell dataset and proposes a network named Binary Tree-like Network with Two-path Fusion Attention Feature (BTTFA). The simplified diagram of BTTFA is similar to a binary tree structure and combines ResNeXt's last four layers of information by integrating adjacent pairs of layers each time; therefore, the information of multilayers can be fully integrated, and the information lost by the pooling layers can be recovered. BTTFA also applies a two-path fusion attention to selectively utilize information close to the root nodes, thereby taking full advantage of low-level detail and high-level semantic information and selectively emphasizing important features while suppressing less useful ones. Meanwhile, at some nodes of the binary tree-like network, focal loss is imposed to calculate the loss between the ground truth and the feature map during the training process. Experiments demonstrate that BTTFA performs better than the existing technology on our dataset and another public dataset.