Noise Reconstruction and Removal Network: A New Way to Denoise FIB-SEM Images

Katya Giannios, Abhishek Chaurasia, Bambi DeLaRosa, Guillaume THIBAULT, Jessica L. Riesterer, Erin S Stempinski, Terence P Lo, Joe W Gray

Keywords: Neural Network, CNN, LSTM, Unsupervised learning, Denoising, FIB-SEM

Abstract: Recent advances in Focused Ion Beam-Scanning Electron Microscopy (FIB-SEM) allow the imaging and analysis of cellular ultrastructure at nanoscale resolution, but the collection of labels and/or noise-free data sets has several challenges, often immutable. Reasons range from time consuming manual annotations, requiring highly trained specialists, to introducing imaging artifacts from the prolonged scanning during acquisition. We propose a fully unsupervised Noise Reconstruction and Removal Network for denoising scanning electron microscopy images. The architecture, inspired by gated recurrent units, reconstructs and removes the noise by synthesizing the sequential data. At the same time, the fully unsupervised training guides the network in distinguishing true signal from noise and gives comparable/even better results than supervised approaches on 3D electron microscopy data sets. We provide detailed performance analysis using numerical as well as empirical metrics.

One-sentence Summary: A fully unsupervised network for denoising a sequence of scanning electron microscopy images.