Integration of U-Net, ResU-Net and DeepLab Architectures with Intersection Over Union metric for Cells Nuclei Image Segmentation

Abstract

Identifying cells nuclei is the starting point for most biomedical analyzes, because cells contain a nucleus filled with DNA, automating the detection of cells nuclei will speed up disease research and help find cures. The goal is to integrate Neural network architectures such as U-Net, ResU-Net and DeepLab with the Intersection Over Union (IoU) quality measure for the segmentation of images of cells nuclei. The dataset is made up of images of cells nuclei found in the Data Science Bowl 2018 which were predicted and successfully segmented by the U-Net architecture. The U-Net IoU metric was identified as having good values, approximately 0.90, which indicates a good response to image modifications, ResU-Net has achieved satisfactory results, but without surpassing U-Net performance. On the other hand, DeepLab did not show satisfactory results which can be improved through network modifications.