VolkCell Guide
A practical guide to VolkCell.
A practical field guide for running analyses, reviewing results, correcting masks, training custom models, managing experiments, and exporting data from VolkCell.
Fastest path
- 1. Open Analysis.
- 2. Upload representative microscopy images.
- 3. Pick the closest protocol/model settings.
- 4. Review counts, masks, CSVs, and overlays.
- 5. If needed, correct masks and train a custom model.
01 · Overview
Use Overview as your workspace summary
What it shows
- Your signed-in workspace and recent activity.
- Counts for active jobs, completed jobs, and available models.
- Quick-access experiment cards for the workflows enabled for your account.
What to click
- Analyse images (Run your first analysis on a new account) for a simple upload-and-count job.
- Quick access cards when you want protocol-specific experiment structure.
- Recent activity to reopen a prior job and download results.
01b · Pipeline agent
Describe your assay and let the agent configure everything
Every experiment page has a "Describe your assay" panel at the top. Instead of manually choosing analysis type, ROI, sensitivity, and ordering, you can describe your experiment in plain language and the agent will configure the pipeline for you.
How it works
- Click "Describe your assay" to expand the agent panel.
- Type a natural-language description of your experiment (e.g. "Adherent HEK293 cells in a 96-well plate, count and confluency").
- The agent looks at sample images from your experiment, considers past run history, and selects the optimal settings.
- Configuration chips appear showing what the agent chose, and the form fields update automatically with a brief highlight.
Tips
- Be specific: mention cell type, plate format, imaging modality, and what you want to measure.
- You can refine: if the agent picks the wrong settings, reply with a correction (e.g. "Actually use scratch assay mode") and it will adjust.
- Sensitivity is automatic — the agent picks the right detection threshold based on your images. You don't need to choose it manually.
- You can always override the agent's choices by changing the form fields directly.
02 · Analysis
Run a quick cell-count analysis
Choose the right images
Use representative brightfield or phase-contrast microscopy images. Keep acquisition settings consistent within a job where possible. Avoid mixing very different objectives, stains, or protocols unless you are intentionally stress-testing a model.
Select the closest workflow settings
Pick adherent count + confluency for monolayer counting, suspension for floating cells, spheroid for 3D cultures, or scratch assay for wound closure. Alternatively, use the pipeline agent — describe your assay in plain language and it configures everything automatically.
Set calibration when scale matters
If you need density or area-normalized outputs, enter the calibration metadata or rely on extracted scale metadata when available. If you only need raw cell counts, calibration is optional. Detection sensitivity is chosen automatically based on your images — you do not need to set it manually.
Submit and wait for processing
Small jobs usually finish quickly. Larger image sets may sit in pending/processing while queue capacity, GPU startup, or file preparation completes. You can leave the page and reopen the job from recent activity.
03 · Results & corrections
Review outputs before trusting them
Counts
Start with totals and per-image counts. Look for obvious outliers before exporting or training from the run.
Masks & overlays
Use overlays to spot missed cells, merged clusters, debris, or scale-bar artifacts. The visual check is the quality gate.
Corrections
Edit masks when the model is wrong, save corrections, then use those corrected images as training signal.
- Correct enough diverse examples to cover normal images, crowded areas, faint cells, debris, and edge cases.
- Do not train from sloppy corrections: the model learns exactly what you save.
- Keep raw outputs and corrected outputs separate when reporting validation results.
04 · Training
Train a protocol-specific model when the default is not enough
Training adapts VolkCell to your cells, microscope, and segmentation preferences. It should follow inspection and correction, not replace it.
Recommended flow
- Run an analysis on representative images.
- Open the job and correct masks.
- Start a training session from the corrected job.
- Train, test on held-out images, inspect benchmark metrics.
- Save/promote the model only when validation is acceptable.
When not to train
- The default masks already match your counting rule.
- You only have one or two corrected images.
- Your corrections are inconsistent between reviewers.
- You need a validated model today but have no holdout images.
05 · Experiments
Use Experiments for structured scientific runs
Experiments group datasets, protocols, model runs, galleries, benchmark reports, and notes so analyses are traceable rather than one-off files. Each experiment has a pipeline agent that can configure settings from a plain-language description of your assay.
Core workflow
Adherent count + confluency
Cell count and area coverage for adherent monolayer cultures. Supports per-well ROI detection and edge-margin exclusion.
Core workflow
Suspension counting
Cells/mL concentration from suspension culture images with optional trypan blue live/dead viability scoring.
Specialist workflow
Spheroid analysis
Per-object size, area, and roundness measurement for 3D spheroid and organoid cultures. Tracks diameter trends over time.
Specialist workflow
Scratch assay
Wound-closure measurement over time from matched scratch-assay images. Reports open wound area percentage at each timepoint.
Advanced
Morphology
Image-level morphology notes and tags alongside counts. You define what to flag — no universal scoring rubric.
06 · Models
Choose models conservatively
Built-in models are the safest starting point for exploratory analysis and first-pass counting.
Experimental custom models are useful for iteration, but treat them as unvalidated until they perform well on held-out images.
Validated custom models should be tied to a protocol, imaging setup, and benchmark report. Do not assume a model validated for one cell type or objective transfers to another.
07 · Exports
Export the right artifact for the job
CSV exports
Use CSVs for count tables, summaries, downstream statistics, and sharing compact results with collaborators.
Mask/overlay bundles
Use bundles when you need to audit segmentation quality, preserve visual evidence, or hand off correction tasks.
Experiment reports
Use reports for protocol-level summaries, benchmark interpretation, and customer-facing validation narratives.
08 · Troubleshooting
Common problems and what to do
- Counts look too high.
- Check overlays for debris, split cells, or scale-bar/text artifacts being counted. Try a more appropriate analysis mode or correct masks before training.
- Counts look too low.
- Look for faint cells, low contrast, crowded clusters, or mismatched model/protocol settings. Add representative corrections and retrain if the pattern repeats.
- Job appears stuck.
- Large jobs and GPU-backed training can pause while files prepare or workers start. Refresh the job from recent activity. If it remains stalled, contact support with the job ID.
- Model is not in the picker.
- Confirm you are signed in with the same email used to train/save it, and that the model was promoted after validation.
- The agent picked the wrong settings.
- Reply in the agent chat with a correction (e.g. "Use spheroid mode instead" or "Switch to hard sensitivity"). The agent remembers your conversation and adjusts. You can also override any setting directly in the form.
- Sensitivity seems off — too many or too few detections.
- The system chooses detection sensitivity automatically based on your images. If results are consistently wrong, tell the agent "use hard sensitivity" for dense or low-contrast images, or "use normal sensitivity" for cleaner fields. The agent adjusts without exposing the technical setting.