Potential issue with gold standards (or fifty shades of myelin)

I was experimenting with a model and I used data from the folder SEM_3c_512 (Duke) and I found that the gold standards contained a range of pixel intensities that weren’t one of the three classes [background, myelin, axon]. For instance, if I just want the myelin content, I threshold for the myelin pixel intensity and it creates a ground truth like this one below:

Where we can see the noisy pixels on many places. Please note that GitHub is resampling these images, so they look less bad then they’re in reality.

When we look on the original gold standards, we can see that there are indeed many different shades representing the classes:

This might be an effect of resampling of the gold standards, that might be creating interpolation on the gold standard and therefore messing up with the class definitions. This may be a problem for AxonDeepSeg if the training is being done without taking into consideration these issues, but I’m not sure how these gold standars are being handled today on AxonDeepSeg. I’m not sure if the data used for training is also from this same folder on Duke.

[jcohenadad]

1. We should fix the gold standard at the first place
2. The answer is in the raw data (ie: the image)

[vasudev-sharma]

What do you think @mathieuboudreau / @jcohenadad is it a problem or is it fine ?

it is a problem

So, I have 2 questions in this regard : -

1. Should we handle it in Dataset building(from raw datasets to patches) or during training ?
2. Should we replace this value either by myelin or by axon values else should go we with different approach ?