Mapping the Connectome: Reducing Variability in Convolutional Neural Network Predictions using Post-processing Techniques

Abstract

Convolutional neural networks have shown to be effective models for performing automaticsegmentation of neuronal structures in electron microscopy (EM) images. However, trainingneural networks results in a slightly different model each time, because each training rundepends on nondeterministic factors such as weight initialization. We trained multiple large,deep convolutional neural networks to perform image segmentation on 2D and 3D images, andfound large variability in the outputted prediction of each run. We reduced the variability across different training runs of the same network architecture by applying post-processing techniques. We found that by combining affinity map predictions and performing mean affinityagglomeration, we were able to achieve better results. Using these techniques, we were able toachieve a Rand score of 0.976 and VI score of 0.984 on the ISBI 2012 challenge, and a Randscore 0.9406 and VI score of 0.984 on the CREMI 2016 challenge.