Multi-dimensional gated recurrent units for the segmentation of biomedical 3D-data

Abstract

We present a supervised deep learning method to automatically segment 3Dvolumes of biomedical image data.The presentedmethod takes advantage of a neural network with the main layers consisting of multi-dimensional gated recurrent units.We apply an on-the-fly data augmentation technique which allows for accurate estimations without the need for either a huge amount of training data or advanced data pre-or postprocessing. We show that our method performs amongst the leading techniques on a popular brain segmentation challenge dataset in terms of speed, accuracy and memory efficiency. We describe in detail advantages over a similar method which uses the well-established long shortterm memory.