Res-CR-Net, a residual network with a novel architecture optimized for the semantic segmentation of microscopy images

Abstract

Deep neural networks (DNN) have been widely used to carry out segmentation tasks in both electron microscopy (EM) and light/fluorescence microscopy (LM/FM). Most DNNs developed for this purpose are based on some variation of the encoder-decoder U-Net architecture. Here we show how Res-CR-Net, a new type of fully convolutional neural network that does not adopt a U-Net architecture, excels at segmentation tasks traditionally considered very hard, like recognizing the contours of nuclei, cytoplasm and mitochondria in densely packed cells in either EM or LM/FM images.