AUTHOR=Xue Xi , Kamata Sei-Ichiro 

TITLE=Contextual Mixing Feature Unet for Multi-Organ Nuclei Segmentation

JOURNAL=Frontiers in Signal Processing

VOLUME=Volume 2 - 2022

YEAR=2022

URL=https://www.frontiersin.org/journals/signal-processing/articles/10.3389/frsip.2022.833433

DOI=10.3389/frsip.2022.833433

ISSN=2673-8198

ABSTRACT=Nuclei segmentation is fundamental and crucial for analyzing histopathological images. Generally, a pathological image contains tens of thousands of nuclei and exists clustered nuclei so it is difficult in separating each nucleus accurately. Challenges against blur boundaries, inconsistent staining, overlapping regions have adverse effects on segmentation performance. Besides, nuclei from various organs appear quite different in shape and size, which may lead to the problems of over segmentation and under segmentation. In order to capture each nucleus on different organs precisely, characteristics about both nuclei and boundaries are of equal importance. Thus, in this paper, we propose a contextual mixing feature Unet (CMF-Unet), which utilizes two parallel branches, nuclei segmentation branch and boundary extraction branch, and mixes complementary feature maps from two branches to obtain rich and integrated contextual features. To ensure good segmentation performance, a multi-scale kernel weighted module (MKWM) and a dense mixing feature module (DMFM) are designed. MKWM, used in both nuclei segmentation branch and boundary extraction branch, contains a multi-scale kernel block to fully exploit characteristics of images and a weight block to assign more weights on important areas so that the network can extract discriminative information efficiently. To fuse more beneficial information and get integrated feature maps, DMFM mixes the feature maps produced by MKWM from two branches to gather both nuclei information and boundary information and links the feature maps in a densely-connected way. Because the feature maps produced by MKWM and DMFM are both sent into the decoder part, the segmentation performance can be enhanced effectively. We test the proposed method on multi-organ nuclei segmentation (MoNuSeg) dataset. Experiments show that the proposed method not only performs well on nuclei segmentation but also has a good generalization ability on different organs.