Using the Training View
The Training view is where you create training samples, train a classifier for one staining, and review that classifier's output.
Choose the staining and slice
Focus on one staining at a time, because each staining has its own independent classifier.
Move through slices to collect variety when the staining looks different across slices.
In practice, using training samples from around 3 to 5 visually different slices is often a good starting point.
This is especially useful when the staining pattern, intensity, or background changes between slices.
Understand the viewer
In most cases, you will work mainly in the staining view, because that image shows whether a cell is expressing the stained protein.
DAPI is mainly a reference.
Unclassified cells appear blue.
Classified cells appear green for ON and red for OFF.
The colors you see represent the output of the currently selected model.
Label single cells
When you label a cell, it is added as a training sample.
Labeled cells are shown with a green or red circle, depending on whether they are ON or OFF.
This makes it easy to see which cells are already part of the training set.
You can also relabel or remove samples when needed.
Select and label many cells
Rectangle selection is the main workflow for collecting samples quickly.
When you draw a selection, Neuro QP does not add every cell in the area. It randomly samples cells from that region.
This is useful because random sampling gives a better representation of the real data.
In practice, a good workflow is to select a region with around 20 to 50 cells inside an interesting region and label them as one class, for example OFF.
Then correct the sampled cells that are wrong, for example by selecting all cells that should be ON.
This works well because it is fast and still keeps the labels representative of the real data.
Usually, you should not try to label every visible cell in an area.
That tends to add more labeling errors and is usually worse than focusing on variety and quality.
As a rule of thumb, do not add much more than 50 samples at a time before checking them.
Use shortcuts to work faster
1for ON2for OFFDeletefor removeDfor DAPI- arrow keys for slice navigation
Escto clear selection
Use the Cells panel
The Cells panel shows the current classifier output for the slice.
It also shows how many training samples are selected on the current slice and how many you have in total.
This helps you keep track of dataset size while you work.
If you already have more than around 2000 samples, the priority is usually improving sample quality and variety rather than adding more samples.
Train a model
Training uses the current training samples for the current staining.
You need at least 20 samples before Neuro QP allows training.
In practice, it is usually a good idea to collect around 200 samples before training the first model.
Each model is trained from scratch using the current training samples.
It does not continue from a previous staining-specific model or reuse previous training results as the starting point.
After training, the newly trained model is available in the model list.
Review the output
After training, the view updates with the classification results for the newly trained classifier.
The colored cells in the viewer show the output of the model that is currently selected.
You can switch between trained models to compare them.