How Training Works

Neuro QP training learns to classify detected cells for one staining as ON or OFF.

ON means the cell is expressing the protein represented by that staining.

OFF means the cell is not expressing that protein.

For each labeled sample, Neuro QP uses the image around that cell as training input.

The cells are already detected. Training does not find cells. It learns how to classify them for that staining.

Each staining has its own classifier

Each staining is trained independently.

A model trained for one staining is not the model for another staining.

You need to train a separate classifier for each staining you want to analyze.

This matters because different stainings can have very different signal patterns and different classification difficulty.

The ON/OFF classification directly affects the counts used in the analysis.

If classification is weak, the final quantification can be misleading even if the rest of the pipeline is otherwise working well.

Good training improves how trustworthy the biological interpretation is.

Neuro QP takes the labeled cell examples and uses the image around each labeled cell for training.

The model learns which visual patterns are typical for ON cells and which are typical for OFF cells for that staining.

The result is a staining-specific classifier that can then be applied across slices.

A good training set is clear, consistent, varied, and balanced enough to represent the real data.

It should include both positive and negative examples.

In practice, a few hundred samples are often enough to get a useful model.

Harder classification tasks may need more, up to around 2000 samples.

Easier tasks usually need fewer samples.

Training cannot fix poor image quality.

Training cannot fix wrong registration.

Training cannot fix inconsistent labeling.

If the underlying data or labels are unreliable, the trained model will also be unreliable.

In practice, training is usually not a one-step process.

The normal workflow is to label samples, train a model, inspect the results, add better samples, and train again.

This is how users usually reach a strong model efficiently.

The goal is not only to get a high training score.

The real goal is to get classification results that are trustworthy for the actual analysis.

The most important quality control is still checking the output visually on real slices.