Guidelines for choosing min and max sigma for effective training and analysis in WEKA trainable segmentation

Hello all, I am trying to train a model for automatic cell segmentation of immunostained mouse neurons. These neurons are roughly circular and 10 pixels across at the current magnification. I can provide specific examples, but for the moment I am looking for general advice.

I am trying to use the WEKA Trainable Segementation program which is working great on the whole, but I am having trouble choosing the sigma range. I was able to find guidance on the purpose of the various features that can be selected for training, but no guidelines on the range of sigma used for the training and analysis.

Due to the machine I am working on, processing time and memory use are both factors in my research, so minimizing the sigma range would be great, and I am afraid I’m new enough to the field of image analysis that I don’t fully understand how the sigma range translates to training with various features beyond that generally it is the size of the filter applied. But, how does one determine the optimal size range to minimize processing load and the effect of noise. What criteria do other labs use to determine the sigma range for the training? What exactly are the units of sigma in relation to the image?

Thank you all for your time,
Theo Kataras

Hello @tkataras

In general you could think on the sigma as the radius of the window around each pixel that is taken into account to classify it. In other words, it would be the radius of the field of view that is used to classify each pixel.


1 Like

Thank you for the prompt response, that helps a lot. To check, would that mean setting a min sigma of 5 would cause there to be a 5 pixel radius not considered around the pixel being classified?

With a minimum sigma of 5, details of the image smaller than that radius would be most likely ignored, yes.