Dual-stream CNN: Joint Edge and Object for Nucleus Segmentation

Anonymous

Keywords: Nucleus segmentation, dual-stream CNN

Abstract: Nucleus segmentation is a significant task that can contribute greatly to re-ducing the time to develop and validate visual biomarkers for new digital pathology datasets. There are large numbers of clusters of crowed nuclei that mainly causes poor performance of deep learning method. This segmentation task normally can be divided into two sub-tasks: (1) edge detection, and (2) object segmentation. However, the proposed methods before have two stag-es: (1) do edge detection and object segmentation respectively, and (2) fuse the edge and object results from stage one. In this paper, we proposed a Dual-stream CNN, which joints edge and object in an end-to-end segmentation network and do feature fusion at extractor by using a cross-stitch module. We use two same U-nets for edge detection and object segmentation and design a cross-stitch module at each stage of both U-nets to combine edge and object features. We validate our method using the Multi-organ Nucleus Segmentation (MoNuSeg) Challenge. Our Dual-stream CNN achieves the state-of-the-art performance on this dataset.