Naive Segmentation

This approach was done just to try for a simple alternative when we had difficulties in implementing other approaches. It includes finding some relations between the pixel intensities of neuron and non-neuron pixels. Idea is to mask the neuron pixels from other pixels and then extract the masked segments.

Images from the dataset are read as grayscale using scipy.misc.imread(). A matrix of dimensions [512x512, N] is built where N is the number of images in the dataset. In different tries, different matrices with mean of images, means of distributions, median of distributions were constructed to analyze a relation.

Issue faced: We couldn't find a decent relation between the distribution of neuron and non-neuron pixels as the pixel values vary across all over the domain [0, 255]. The mean distribution of neuron pixels lies majorly in range [70, 220] and that of non-neuron pixels are majorly in range [0-30] and [240-255] but that doesn't rule out the fact there are a lot of non-neurons with pixel values in range [70-220] making it difficult to determine if pixel belong to neuron or not. Though after trying out some maskings, large blobs which include surrounding non-neuron pixels of closer pixel values. Thus a better approach like neural network or some image processing techniques have marked importance to make good neuron segmentation.