CNAC-Seg: Effective segmentation for cervical nuclei in adherent cells and clusters via exploring gaps of receptive fields

Abstract

Computer-aided diagnosis technology is the key for scaling up cervical cancer screening. Existing techniques can clearly segment individual nuclei but are difficult to achieve satisfactory results on adherent nuclei. Thus, we propose an effective method(CNAC-Seg) to accurately segment adherent cervical nuclei via exploring gaps of receptive fields. Specifically, CNAC-Seg devises Enhanced Multi-scale Receptive Fields(EMRF) module, which extends multi-scale receptive fields with large gaps by capturing the lost spatial information and reweighting the imbalanced weights. Furthermore, Feature Refinement Module(FRM) creatively concentrates on extracting intra-level features with small-gap receptive fields to obtain more delicate semantic representations, thereby obtaining refined adherent cervical nuclei’s boundaries. Additionally, we incorporate an extra supervision signal which modifies the features learned from the backbone for two different optimization directions of segmentation. By collaboratively learning large-gap and small-gap receptive fields, CNAC-Seg can clearly segment adherent nuclei to help doctors better diagnose the extent of cancer lesions. Extensive experiments demonstrate the superiority of CNAC-Seg over the state-of-the-art methods, achieving the best IoU results of 0.5234, 0.8585, and 0.9034, the best Dice results of 0.6872, 0.9239, and 0.9492, the best PA results of 0.6293, 0.9758, and 0.9434 in Our Dataset, ISBI, and Herlev respectively.