Augmentation Parameters.csv
Advanced user only, you can safely skips this part of wiki if it's the first time you are using VST!
This tool use two csv files for storing the parameters of image augmentations. One for isotropic resolution dataset while the other one for anisotropic (default).
A technique to expand the variety of dataset.
When number (and size) of images to train the network are few, the network could over-fit on your training images, and have bad performance on images it didn't see before.
By apply various techniques including crop, flip, rotation and so on, you can generate many variations of your training images, which helps to overcome the over-fit issue.
Most of the time you don't need to modify the augmentation.csv file. VST included Augmentation Parameters Anisotropic.csv and Augmentation Parameters Isotropic should be sufficient.
Rotate the image at xy/xz/yz plane respectively, with a random angle between Low Bound and High Bound.
Note if your images are anisotropic (depth resolution is different to xy resolution), you should not enable xz/yz plane rotation.
Each cropped patch can be randomly scaled with a scale factor between Low Bound and High Bound.
Note this won't affect the size of each cropped patches.
Lower than 1 is zoom in, while higher than one is zoom out.
Crops the input tensors to a smaller version, and replicate pad the lost region. Do not turn this on, this is just for testing purpose.
Self Explanatory.
Simulate lower imaging resolution by down-sampling using nearest-neighbor interpolation and then up-sampling using cubic interpolation.
Add Gaussian noise with a mean of 0 and a std of 1, with a strength between low bound and high bound.
Apply 3D Gaussian blur. Its low bound and high bound refers to the range of sigma of the blur. While its value is the kernel size, higher than 3 could slow down augmentation process significantly.
Self Explanatory. Do note that since the images are normalised to have a mean of 0 and std of 1, changing contrast is actually just multiplying the image by a value; changing gamma is first scale image to be between 0 and 1, apply gamma adjustment, then scale back; changing brightness is just add value to the entire image.
Apply gamma or contrast adjustment to with a gradient pattern. From Low Bound at one (random) side of the image to High Bound at the opposite side of the image.
Used to simulate uneven staining throughout the sample.
Add salt and pepper noise. The proportion of pixels affected are between Low Bound and High Bound.
Blur the Label and the contour Label a little bit. Please do not enable them as they have shown to be ineffective.