Shulei Chang Tingting Yang Bowen Yin Jiayi Zhang Liang Ma Yanhui Ding Xiaodan Sui ^*

• Shandong Normal University, Jinan, China

Automation of nucleus segmentation from histopathology images facilitates efficient and accurate identification of the tissue structure and nuclear morphology. However, nuclear segmentation remains a challenge due to the diversity of nuclear margins. Existing methods have not used flexible sampling and balanced semantic features to segment nuclear pathology images. We propose a new method based on a convolutional neural network (CNN) called the deformable multi-level feature network (DMFNet) to enhance nucleus segmentation performance on pathology images. We first divide the DMFNet into two levels, namely feature level and mask level. At the feature level, we enhance the transformation modeling capability of the network and enhance the original features with multi-level features. In the mask level, we use a one-stage framework to make the model simpler. The proposed method achieved the best performance on the MoNuSeg 2018 dataset, with its mAP and mAR approaching 37.8% and 47.4%, respectively, when comed to numerous state-of-the-art methods. In this study, we proposed a new model, called the DMFNet, for nuclear segmentation from digital pathology images of different organs. This study analyzed the process of cell nucleus segmentation, and improved the feature extraction, fusion, and mask generation processes. The proposed DMFNet could effectively combine a DCN, BFP, and SOLO to improve the performance of the segmented network.