Gated CNNs for Nuclei Segmentation in H&E Breast Images

Nuclei segmentation using deep learning has been achieving high accuracy using U-Net and variants, but a remaining challenge is distinguishing touching and overlapping cells. In this work, we propose using gated CNN (GCNN) networks to obtain sharper predictions around object boundaries and improve nuclei segmentation performance. The method is evaluated in over 1000 multicentre diverse H&E breast cancer images from three databases and compared to baseline U-Net and R2U-Net.