Dense-U-Net: densely connected convolutional network for semantic segmentation with a small number of samples

Abstract

The main contribution of this work is the proposal of a densely connected convolutional network for semantic segmentation, which strengthens utilization of features and improves segmentation results even with limited training samples. To achieve this, we combine the U-Net network and our resulting system is called Dense-U-Net. Compared to traditional convolutional networks such as U-Net, there are additional concatenation layers between each pair of convolutional layers which have the same size of outputs in our Dense-U-Net, each layer can get the feature-maps of all its preceding layers as inputs while its feature-maps can be passed to all subsequent layers, and a higher segmentation quality can be achieved without a need for increasing the volume of datasets finally. We evaluate our proposed architecture by segmentation accuracy, foreground-restricted rand scoring after border thinning V^Rand and foreground-restricted information theoretic scoring after border thinning V^Info at the same time, and the results are shown on three different segmentation tasks: ISBI challenge 2012 for segmentation of neuronal structures in electron microscopic stacks, ISBI cell tracking challenge 2014(Glioblastoma-astrocytoma U373 cells) and 2015(HeLa cells), our Dense-U-Net achieves better results than U-Net and several other state-of-the-art networks on all tasks.