An Improved Breast Cancer Nuclei Segmentation Method Based on UNet++

Abstract

Nuclei segmentation plays an important role in medical image analysis but it is also a challenging area due to the tiny size of nuclei especially for breast cancer nuclei. To address these challenges, in this paper we present an improved UNet++ architecture, a more powerful architecture for nuclei segmentation. The original UNet++, which is an encoder-decoder architecture with a series of nested and dense skip pathways, is used as the framework in our work. The main reason for the increase in ability is that the Inception-ResNet-V2 network is added as backbone, which is a very deep network with brilliant performance in object detection. We have evaluated our improved UNet++ in comparison with UNet and the original UNet++ architectures in breast cancer nuclei segmentation dataset. The experiments demonstrate that our improved UNet++ is superior to U-Net and the original U-Net++.