Multiscale Dilated UNet for Segmentation of Multi-Organ Nuclei in Digital Histology Images

Abstract

Millions of deaths occurs every year due to various kinds of cancer. Late diagnosis and no proper treatment planning are the main contributing factors of these deaths. Tissue slides are commonly used for tumor assessment by extracting bio-markers from the biopsies. These bio-markers are then further used for cancer diagnosis. Digitized tissue slides contain multi gigapixels which is why automatic tumor segmentation methods have been developed. However, these methods fail to delineate accurate boundaries as well as are unable to detect objects at multiple scales. Therefore to eradicate this problem we have proposed Multi-scale Dilated U-Net (MD-UNet) which performs feature extraction at multiple scales and delineate accurate boundaries. MD-UNet is trained on 5 Nuclei Segmentation datasets each belonging to different organ of human body. The proposed model outperforms DeepLab v3+, SegNet, U-Net and U-Net++ on all the 5 Nuclei Segmentation datasets.