Interpreting Training Results

After training, NeuroQP shows classifier output in the viewer and short hints about the training run.

Use that information to decide what to improve next.

Start with visual review

The most important quality control is checking real slices visually.

Look at whether ON/OFF calls make sense for the selected workflow.

For cell counting and classification projects, check whether marker-positive and marker-negative cells are separated correctly.

For independent-detection projects, check whether accepted marker-positive cells remain visible and false-positive candidates are rejected.

Use hints as guidance

Training hints can point toward likely weaknesses in the current model.

They should not be treated as the final definition of model quality.

Accuracy is only one hint

Accuracy or F1 can be useful, but the real goal is classifier behavior that is trustworthy on the actual slices used in analysis.

Scores can go down when you add harder samples even while the model becomes more useful.

Decide the next step

If the output is wrong in a specific way, add samples that target that failure mode.

Use the hints and visual review together.

When the model is good enough

A model is good enough when the results look trustworthy on representative slices.

The goal is not to maximize a number. The goal is reliable analysis output.