How to train your own model

Dashboard Start training →

Training a custom model adapts VolkCell to your cells and imaging conditions. The Training page guides you through five stages: Upload → Correct → Train → Test (optional) → Save. You can leave the page at any point and continue the same session later.

First, check that training fits your measurement

Custom training teaches the model which pixels belong to each cell. It can improve cell-outline and count results. It does not teach tube-network measurements (length, branches and junctions) or a live/dead viability threshold.

Walkthrough video

Coming soon

Video coming soon

The walkthrough will appear here

The five stages

The stage bar on the Training page shows where you are. Each stage has a clear "what you do" and "what happens automatically". Read the second one only if you want to know the why.

  1. 1

    Upload your images

    Pick 5–15 representative images covering the densities, lighting and cell shapes you actually deal with. More variety is more useful than more near-identical fields. At least one usable corrected image is required before a trial run can start.

    If you already have a completed analysis, choose Use previous job instead of uploading the same images again.

    What happens automatically: VolkCell checks each file and prepares initial cell outlines for you to correct. You can leave the page while this runs and return with Continue training.

  2. 2

    Correct the outlines

    Open each image in the editor and fix what the starting outlines got wrong: add missed cells, remove extra objects and separate touching cells. Save the corrected mask, then return to the Training page.

    Several saves on the same image still count as one corrected image. Corrections with an unreadable or empty mask, or a mask that does not match the image size, are flagged and cannot be used for training.

    What happens automatically: Every usable saved correction in the session becomes a labelled example for training. You do not need to approve each correction separately before training.

  3. 3

    Train

    When at least one usable correction is saved, the training step unlocks. Give the model a unique name and start the run. A set with around 5 corrected images and 20 corrected objects is more useful than a single image, though any small run is still a trial.

    Training continues in the background. You can close the tab and return with Continue training.

    What happens automatically: VolkCell builds the training set from the session's usable corrections. With five or more corrected images, a small portion is kept out of fitting as an automatic check. That split is at image level, so fields from the same acquisition can still be closely related; it is not independent validation.

  4. 4

    Test on fresh images (optional)

    Upload one or more images that were not used for training and compare the outlines with the image. The overlay is the first check: if it looks wrong, the model is not ready regardless of any score.

    Review the result

    If the outlines are right, continue to Save. If not, correct the uploaded test image and retrain with it so the model sees that failure case. Repeating this is normal.

    Testing is optional. You can save before testing with Save as experimental, but testing on fresh images is recommended and saving does not validate the model.

    What happens automatically: The trained model runs on the fresh images through the same path as your normal analyses. The validation benchmark compares it with the original analysis on corrected source images; many of those images were also used for training, so the scores describe performance on those examples rather than independent accuracy. Scores from independently corrected test images are stronger but still limited to those images. How to read the benchmark.

  5. 5

    Save your model

    Choose a unique name and save. The model is added to the model picker in Analysis for your account and can be selected manually. New models start as experimental: saving does not validate accuracy or enable automatic use.

    Testing lets you review the model on fresh images before saving. If you saved without a test, review it on independent images before relying on its results.

    Continue training: if you leave before saving, the Training page shows a Continue training banner for your most recent unfinished session. Use it to return without uploading again. Uploading new images starts a separate session.

    What happens automatically: The completed model is added to your model list and kept for later analyses. Saving does not enable automatic use.

Starting from a previous analysis

If you already ran an analysis, you do not need to upload the same images again. On Upload choose Use previous job and select a completed analysis; the session opens at Correct with those images.

From a completed job

On the Upload stage, choose Use previous job and select the analysis. Its images and saved corrections are reused, so you can correct what needs work and go straight to training.

Saved models work across analyses

A saved model belongs to your account and can be selected manually in the Analysis model picker for later jobs. It is not tied to one experiment.

Tips for a strong model

Diversity beats volume

Images covering low / medium / high density and clean / messy fields are generally more useful than many near-identical ones.

Correct the failures

When iterating, prioritise correcting the images where the previous model did worst. That is where the model has the most to learn.

Be ruthless about edges

Touching cells are the hardest case. A few well-corrected clusters are valuable training examples, so do not skip them.

Don't over-correct

If a cell is genuinely ambiguous, leave it alone. Forcing a label you'd disagree with tomorrow makes the model worse, not better.

How to read the validation benchmark

Start with the fresh test image overlay. If that overlay looks wrong, the model is not ready no matter what the numbers say. The benchmark scores add precision; they do not replace judgment.

F1 balances misses and false positives. Higher is better.
IoU measures overlap with the corrected mask. Higher is better.
Precision tells you how many of the objects or pixels the model marked were actually right. Higher is better.
Recall tells you how many real cells or pixels were found. Higher is better.
Count error shows how far the predicted count is from the corrected count. Lower is better.
Important: these scores benchmark the model on corrected source images from the analysis you trained on. Many of those images were also used for training, so the scores are useful for measuring improvement, but they are not independent validation. A fresh test image has no corrected ground truth until you correct it; when fresh test images are corrected, their held-out metrics are stronger, but they still only cover those images.

Common questions

How many images do I really need?
One usable corrected image unlocks a trial run. Aim for at least five corrected images and around twenty corrected objects, and add variety where you can. No image count guarantees accuracy; test the result on fresh images.
How long does training take?
It depends on the number and size of your images and the training settings. The Training page shows progress while it runs.
Can I keep using the app while it trains?
Yes. Training runs in the background and does not block normal analysis. You can close the tab and return later with Continue training.
Can I save a model without testing it?
Yes. Use Save as experimental. A newly saved model is experimental, is available for manual selection, and should be reviewed on independent images before routine use.
My new model is worse than the default. What now?
That often means too few corrections, or corrections that disagree with each other. Return to Correct, audit your saved masks, add more diverse images, and retrain. Use Test to compare on fresh images; if the result is still worse, keep using the model you already trust.
Can I delete a model I no longer want?
There is no self-service delete for saved models yet. If a saved model is wrong, contact support so we can review it. Analyses that already used the model keep their recorded results.

Ready to start?

You can come back to this page any time from the training screen.

Start training