Computer Vision News - August 2018

[INFERENCE] Many networks are fully convolutional (with no fully connected layers), thus the resolution of the network output image can be different from that of the input image. This section allows you to configure the output image size that you need. The input image size is NxNxN, output image size is DxDxD. In networks where D is equal to N, no sampling is needed, and thus border will be set to 0. For networks where D<N border should be set to at least (N-D)/2 and the spatial_window_size needs to be set appropriately. [SEGMENTATION] This section defines the parameters for the segmentations the network will perform (type of image, number of segmentations, etc.). E. Gibson, W. Li, C. Sudre, L. Fidon, D. Shakir, G. Wang, Z. Eaton-Rosen, R. Gray, T. Doel, Y. Hu, T. Whyntie, P. Nachev, D. C. Barratt, S. Ourselin , M. J. Cardoso and T. Vercauteren (2017) NiftyNet: a deep-learning platform for medical imaging , CMPB (journal) 20 Tool Computer Vision News NiftyNet sample_per_volume = 32 rotation_angle = (-10.0, 10.0) scaling_percentage = (-10.0, 10.0) lr = 0.01 loss_type = Dice starting_iter = 0 save_every_n = 5 max_iter = 10 max_checkpoints = 20 border = (24, 24, 24) #inference_iter = 10 save_seg_dir = ./output/unet output_interp_order = 0 spatial_window_size = (105, 105, 105) image = T1 label = parcellation output_prob = False num_classes = 160 label_normalisation = True