U-Net##: A Powerful Novel Architecture for Medical Image Segmentation

Abstract

As medical image segmentation has been one of the most widely implemented tasks in deep learning, there have been various solutions proposed for its applications to achieve better results. The encoder-decoder based U-Net architecture and its variants have shown outstanding performance in this field. However, most of these solutions have limited capacity to extract sufficient features from the input images. In this paper, we propose a powerful novel architecture named U-Net##, which consists of multiple overlapping U-Net pathways and has the strategies of sharing feature maps between parallel neural networks, using auxiliary convolutional blocks for additional feature extractions and deep supervision, so that it performs as a boosted U-Net model for medical image segmentation. Our architecture is essentially a combination of encoder-decoder based multiple U-Net pathways which have different depth levels, and all their overlapping feature maps on the same sampling steps share their own feature data with the others by following a specific addition rule. Each network pathway has its own concatenated long skip connections from their encoder to decoder sections, and the final output is obtained with deep supervision method. All these strategies help the model explore much more features effectively and achieve higher accuracy. PyTorch implementation of the U-Net## with step-by-step coding is available here: https://github.com/firatkorkmaz/unetsharpsharp