On Improving 3D U-net Architecture

Abstract

This paper presents a review of various techniques for improving the performance of neural networks on segmentation task using 3D convolutions and voxel grids – we provide comparison of network with and without max pooling, weighting, masking out the segmentation results, and oversampling results for imbalanced training dataset. We also present changes to 3D U-net architecture that give better results than the standard implementation. Although there are many out-performing architectures using different data input, we show, that although the voxel grids that serve as an input to the 3D U-net, have limits to what they can express, they do not reach their full potential.