How Training Works
NeuroQP training learns a staining-specific classifier from positive and negative examples.
The meaning of a positive or negative example depends on the project workflow.
Cell counting and classification projects
In this workflow, cells are detected once from a Cell Detection marker such as DAPI or NeuN.
Marker stainings are then classified against those detected cells.
- A positive example is a detected cell that expresses the selected marker.
- A negative example is a detected cell that does not express the selected marker.
This workflow has a shared cell denominator, so marker-positive proportions are meaningful.
Independent-detection projects
In this workflow, candidate cells are detected directly from each marker staining.
Training teaches NeuroQP which candidates should count as accepted marker-positive cells.
- A positive example is an accepted marker-positive cell.
- A negative example is a rejected candidate, false positive, or background.
A negative example is not a biological marker-negative cell in this workflow. It means the detection should not be counted.
Each staining has its own model selection
Each staining is evaluated with its own classifier model.
A model trained or imported for one staining is not automatically the model for another staining.
You need a selected classifier model for each staining you want to analyze.
How the model learns
NeuroQP takes labeled examples and uses the image around each detected candidate or cell for training.
The result is a staining-specific classifier that can be applied across slices.
What training cannot fix
Training cannot fix poor image quality, wrong registration, missing images, or inconsistent labels.
If the underlying data or labels are unreliable, the trained model will also be unreliable.
Training is iterative
The normal workflow is to label samples, train a model, inspect the output, add targeted samples, and train again.
The most important quality control is visual review on real slices.