2019/06/14 by Robin Brügger, Brügger, Robin, Christian F. Baumgartner +3 · 1 citation
Computer Science · #Advanced Image and Video Retrieval Techniques #Advanced Neural Network Applications #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #FOS: Electrical engineering #Human Pose and Action Recognition #Image and Video Processing (eess.IV) #electronic engineering #information engineering
paper · pdf · doi:10.48550/arxiv.1906.06148
openalex publication_date 2019/06/14 · openalex created_date 2022/07/28 · openalex updated_date 2026/07/28
One of the key drawbacks of 3D convolutional neural networks for segmentation\nis their memory footprint, which necessitates compromises in the network\narchitecture in order to fit into a given memory budget. Motivated by the\nRevNet for image classification, we propose a partially reversible U-Net\narchitecture that reduces memory consumption substantially. The reversible\narchitecture allows us to exactly recover each layer's outputs from the\nsubsequent layer's ones, eliminating the need to store activations for\nbackpropagation. This alleviates the biggest memory bottleneck and enables very\ndeep (theoretically infinitely deep) 3D architectures. On the BraTS challenge\ndataset, we demonstrate substantial memory savings. We further show that the\nfreed memory can be used for processing the whole field-of-view (FOV) instead\nof patches. Increasing network depth led to higher segmentation accuracy while\ngrowing the memory footprint only by a very small fraction, thanks to the\npartially reversible architecture.\n