Computer Vision News - August 2018

memory and transformed into “A, R, S” orientation with voxel size of ( 1.0, 1.0, 1.0 ) with 3 as interpolation order. This is the precise point at which we can see NiftyNet’s customization to suit medical datasets: filename_contains can include a number of different extensions allowing the network to handle a variety of imaging modalities, an option which is not available in general-purpose deep learning systems [SYSTEM] The SYSTEM section specifies to NiftyNet the cuda environment variable, the number of threads to be used and the number of GPUs. The model_dir directory defines where to save intermediate results needed for various computations. [NETWORK] The NETWORK section defines what network to use and the parameters related to network definition (what activation function to use, batch size, type of regularization, etc.). Some other interesting parameters in this section, directly relevant to the medical field, include: window_sampling -- this parameter allows the network to differentially sample, for instance, to sample certain areas more than others; weighted -- to sample a voxel proportionally to the cumulative intensity histogram 18 Tool Computer Vision News NiftyNet path_to_search = ./example_volumes/monomodal_parcellation filename_contains = T1 filename_not_contains = T2 spatial_window_size = (96, 96, 96) pixdim = (1.0, 1.0, 1.0) axcodes=(A, R, S) interp_order = 3 cuda_devices = "" num_threads = 2 num_gpus = 1 model_dir = ./models/model_unet