Settings Explaination
After Starting the GUI, you should see a page like this:
This is the main interface for the tool, where all the configurations will be performed.
Most of the settings are explained in the GUI already, this page only covers those don't have an explanation.
Semantic: The segmentation goal is to predict a class label for every pixel, where the two possible labels are foreground object(s) and background. For example, the foreground object is mitochondria, and the background is everything else, all the mitochondria will get the same label.
Instance: The segmentation goal is to predict a class label for every pixel, and also distinguishes between different instances of the same class. For example, different mitochondria in an image get different label.
This tool does not support segmenting multiple different types of object at once. If there is a need for that, user should try separate their task into multiple smaller tasks.
Training: Train a network using the training dataset.
Validation: Use a Validation Dataset to evaluate the network during training.
Test: Similar to Validation, but only run once after the model has finished training.
Predict: Use a (trained) network to predict the label of the predict dataset.
Note that if the validation is not enabled, the script will use train dataset loss to determine when to reduce the learning rate.
Please read the description on the GUI. Note that setting this too high (>256) could consumes significantly more VRAM and are slower to train.
If this setting is clicked, VST will first load the training images one at a time, and save them as zarr files in the "zarr_preprocessed" folder under your training folder. During training, VST will load from the zarr files. Which uses far less system memory if the training datasets are large.
Tick this if your training data contain large areas that doesn't have any foreground object, or large areas that doesn't have any background object.
Determines the format required for the manual image labels for training.
Fully Labelled: value of 0 is the background, value of 1 is the foreground.
Sparsely Labelled: value of 0 is unlabelled, value of 1 is foreground, value of 2 is background.
Note: Sparsely Labelled data is unsupported in Instance Segmentation Mode.
This setting only work in instance segmentation mode. It determines the width of the contour between objects in the generated contour maps.
It's default to 1, but if the image has very high resolution, it's safe to increase it to 2 for better performance. You should check what the cached contour map in the "generated_contour_maps" folder look like, to determine the suitable value.
Same data:
Width = 1
Width = 2
Same as Training Settings
Since the image can be big, feeding the entire image into the network can have unrealistic VRAM demand. So, it will get cropped into several (potentially) overlapping smaller images. The final results will be stitched together from predictions of these smaller images. This is the height and width of patches the prediction image will be cropped to. In pixels.
The depth of patches the prediction image will be cropped to. In pixels.
The additional gain in height and width of the patches. In pixels. Helps smooth out borders between patches.
The actual height and width of each patch is hw_size + 2 * hw_overlap. Same goes for depth.
Same as above but for depth.
Because instance segmentation is achieved by predicting a contour map, and it's quite often those contour map would have holes in them. Those holes will cause objects not being properly seperated. By using distance transform watershed, this problem can be avoided. However it's much slower, uses much more CPU RAM, and can cause (severe) over-segmentation if the dynamic is not high enough.
Still, it's easier to merge objects than to seperate them, so it's helpful in some cases to tick it.
Path to the file which containing all the parameters for data augmentation.