Neural Networks for Segmenting Neuronal Structures in EM Stacks

A pixel classifier is driving our winning approach to neural structure segmentation in electron microscopy images. Without explicit feature computation, the label of each pixel (membrane or nonmembrane) is predicted from raw pixel values in a square window centered on it. The classifier is a special type of feed-forward neural network trained by plain gradient descent. The input layer maps each window pixel to a neuron. It is followed by a succession of convolutional and maxpooling layers which preserve 2D information and extract features with increasing levels of abstraction. The output layer produces a calibrated probability for each class, and is subjected to very mild post-processing. The approach outperformed all other entries of the competition in all considered metrics: values for rand error, warping error and pixel error were 48 · 10−3, 434 · 10−6 and 60 · 10−3, respectively.