Computer Vision News - August 2018

❏ ApplicationDriver defines the common structure across all applications and is responsible for instantiating the NiftyNet pipeline ❏ applications supported by NiftyNet are: ❏ image segmentation ❏ image regression ❏ image model representation (via auto-encoder applications) ❏ image generation (via auto-encoder and generative adversarial networks (GANs)) ❏ The NiftyNet model zoo includes untrained networks (e.g. U-Net and V-Net for segmentation), and networks trained for specific tasks (e.g. dense_V-Net for multi-organ abdominal CT segmentation, W-Net for brain tumor segmentation and simulator_gan for generating ultrasound images). ❏ The Reader class loads medical-format image files, and applies image- wide preprocessing ❏ Sampler objects generate a sequence of data packets. Each packet contains all the data needed for one independent computation (e.g. one step of gradient descent during training), including images, labels, classifications, noise samples or other data needed for processing ❏ sampling image patches ❏ sampling whole images rescaled to a fixed size and sampling noise ❏ Output ❏ During training: computes a loss and the gradient of the loss for the network output, with respect to the trainable variables, then uses an Optimizer to iteratively train the model. ❏ During inference: aggregates one or more network outputs and performs any necessary postprocessing (e.g. resizing the outputs to the original image size) to generate useful outputs. ❏ The Sampler includes Layer class implementations for data augmentation and normalization -- two means used to compensate for small training data sets in medical image analysis ❏ Data normalization reduces the variability in the data set by transforming inputs to have specified invariant properties, such as fixed intensity histograms or moments (mean and variance) ❏ Data augmentation artificially increases the variability of the training data set, for instance by applying random spatial transformations or adding random image noise ❏ Networks developed and trained outside NiftyNet can be encapsulated within a Network class derived from TrainableLayer . Externally trained weights can be loaded using a restore_initializer . 16 Tool Computer Vision News NiftyNet