Training deep neural networks on large and sparse datasets is still challenging and can require large amounts of computation and memory. In this work, we address the task of performing semantic segmentation on large volumetric data sets, such as CT scans. Our contribution is threefold: 1) We propose a boosted sampling scheme that uses a-posterior error maps, generated throughout training, to focus sampling on difficult regions, resulting in a more informative loss. This results in a significant training speed up and improves learning performance for image segmentation. 2) We propose a novel algorithm for boosting the SGD learning rate schedule by adaptively increasing and lowering the learning rate, avoiding the need for extensive hyperparameter tuning. 3) We show that our method is able to attain new state-of-the-art results on the VISCERAL Anatomy benchmark.
View on arXiv