Computer Vision News - February 2017

10 Computer Vision News We Tried for You Those commands create, train and test a tiny CNN. This verification process takes just a few minutes. You can even test the original CNN presented in the paper. Try it on your own data: Now comes the most interesting part: how to train DeepMedic on your own dataset. All you need to do is to set three configuration files. This process consists of the following steps: defining the model; defining the training; and testing data and settings. 1. Defining the model is done through the modelConfig.cfg file. Let's go over the major parameters of the configuration: a. numberOfOutputClasses – defines the number of classes (categories) the data should be classified into. b. numberOfInputChannels – defines the number of MRI modalities for each subject; in a non-MRI setting this will usually be set to 1. c. segmentsDimTrain – defines the dimensions of the CNN receptive field. d. batchSizeTrain – the number of segments to be included in each batch. e. numberFMsPerLayerNormal – specifies the number of feature Maps to use in each layer. For example, [4,5,6] defines a three-layer network with 4, 5 and 6 features, respectively. Change the numbers in this list to train bigger or smaller networks. f. kernelDimPerLayerNormal – specifies the dimensions of the kernel at each layer. 2. Defining the training data and settings is done through the testConfig.cfg file, the major parameters of this configuration file are: a. channelsTraining – a list of files, each one of them containing a list of images to be used as input. b. gtLabelsTraining – the names to be given to the results (ground truths) of each training image. c. roiMasksTraining – file with a list of file names specifying the Region-of- Interest mask for each training image. 3. Defining the testing data and settings is done through the testConfig.cfg file. The major parameters of this configuration file are: a. channels – a list of files, each one of them containing a list of images to be used as input. b. namesForPredictionsPerCase – the names to be given to the results of each test image. c. roiMasks – file with a list of file names specifying the Region-of-Interest mask for each test image. Once those configurations are appropriately set, you can create, train and test your model with the above three commands (see end of previous page). Tried for You