Deep Volumetric Feature Encoding for Biomedical Images

Abstract

Deep learning research has demonstrated the effectiveness of using pre-trained networks as feature encoders. The large majority of these networks are trained on 2D datasets with millions of samples and diverse classes of information. We demonstrate and evaluate approaches to transferring deep 2D feature spaces to 3D in order to take advantage of these and related resources in the biomedical domain. First, we show how VGG-19 activations can be mapped to a 3D variant of the network (VGG-19-3D). Second, using varied medical decathlon data, we provide a technique for training 3D networks to predict the encodings induced by 3D VGG-19. Lastly, we compare five different 3D networks (one of which is trained only on 3D MRI and another of which is not trained at all) across layers and patch sizes in terms of their ability to identify hippocampal landmark points in 3D MRI data that was not included in their training. We make observations about the performance, recommend different networks and layers and make them publicly available for further evaluation.