Component classification with a convolutional neural network

Components found by the CNMF algorithm can be classified according to their shape into neurons and noise/processes using a Keras pre-trained convolutional neural network classifier. The classifier can be downloaded from the link found in the utilities/cnn_classifier.m file. It accepts 2d image patches of size 50 x 50 that have been obtained by the CNMF algorithm run with half neuron size gSig = 5-7 . The accepted components are those where the output of the CNN network is above a specified threshold value thr. To transform a set of spatial components into a set of 50 x 50 patches you can use the function utilities/extract_patch.m.

Note that the classifier requires MATLAB version 2017b or later, the Neural Networks Toolbox version 2017b or later and the Tensorflow-Keras model importer add on.