The task of three-dimensional data augmentation is computingly coslty and may require a large number of calculations and slow down the process of teaching deep models. You can reduce learning time by optimizing the use of augmentations during training. The article discusses various tools for threedimensional data augmentation: BatchGenerators, Torchio, Rising. The working time of the common three-dimensional augmentations is given: rotation, flip, data crop, blur, application of affine transformations and changes in the size of three-dimensional data on the central processor (CPU) and graphic accelerator (GPU). The advantage of performing augmentation on the GPU in comparison with the CPU is demonstrated, the resulting acceleration is 1-2 orders for most augmentations. A comparison of the time of various augmentations relative to each framework is presented for a more visual choice of augmentations with an emphasis at their time.

Data augmentation, computed tomography, deep learning, benchmarking

Vasiliev Evgeni Pavlovich,  ,  orcid.org/0000-0002-7949-1919,  Lobachevsky State University of Nizhni Novgorod