Zhou Zongwei ; Siddiquee Md Mahfuzur Rahman ; Tajbakhsh Nima ; Liang Jianming

Abstract

In this paper, we present UNet++, a new, more powerful architecture formedical image segmentation. Our architecture is essentially a deeply-supervisedencoder-decoder network where the encoder and decoder sub-networks areconnected through a series of nested, dense skip pathways. The re-designed skippathways aim at reducing the semantic gap between the feature maps of theencoder and decoder sub-networks. We argue that the optimizer would deal withan easier learning task when the feature maps from the decoder and encodernetworks are semantically similar. We have evaluated UNet++ in comparison withU-Net and wide U-Net architectures across multiple medical image segmentationtasks: nodule segmentation in the low-dose CT scans of chest, nucleisegmentation in the microscopy images, liver segmentation in abdominal CTscans, and polyp segmentation in colonoscopy videos. Our experimentsdemonstrate that UNet++ with deep supervision achieves an average IoU gain of3.9 and 3.4 points over U-Net and wide U-Net, respectively.