InvUnet:Inverse the Unet for Nuclear Segmentation in H&E Stained Images

Nuclear segmentation is the starting point for most cancer pathological analysis. Recently,A series of Unet-like deep neural networks have achieved good result in this task. These models are variants of encoder-decoder architecture and can be traced from FCN, which is designed for semantic segmentat...

Full description

Saved in:
Bibliographic Details
Published in:International Conference on Information Science and Technology pp. 251 - 256
Main Authors: Zhang, Lifeng, Li, Bin
Format: Conference Proceeding
Language:English
Published: IEEE 01.08.2020
Subjects:
CNN
H&E stained images
instance segmentation
nuclear segmentation
ISSN:2573-3311
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Nuclear segmentation is the starting point for most cancer pathological analysis. Recently,A series of Unet-like deep neural networks have achieved good result in this task. These models are variants of encoder-decoder architecture and can be traced from FCN, which is designed for semantic segmentation task of natural images. However, compared with natural images, the target regions of H&E images are more salient and have lower entropy. In light of the characteristics of H&E stained images and inspired by the two-stage framework of traditional pure image processing methods, we propose an end-to-end network named InvUnet, which considers nuclear segmentation as a bottom-up stage-wise noise reduction process. The key insight is to take lower-level feature maps of CNN as the output of coarse segmentation, and assume that abstract semantic information from higher layers could guide the denoising process. InvUnet is U-shaped and consists of a contracting path like Unet and a denoising path. Contrary to the top-down structure of Unet's expansive path, the denoising path is bottom-up, starting from the bottom features and combining higher level information to refine the segmentation step-by-step. We conduct experiments on the public avaliable nuclear segmentation dataset. On pixel level segmentation task, InvUnet achieves cleaner segmentation results than Unet with only 1.26M parameters. On instance-level segmentation task, InvUnet combined with weight map for implicit contour learning achieves the best results in separating touching nuclei.
ISSN:2573-3311
DOI:10.1109/ICACI49185.2020.9177722