Measure the biology without hiding the segmentation failures.

This is a new Services page and a fourth card on the Services hub.

A segmentation overlay can look convincing while object counts, boundaries and intensity measurements are wrong. Dense nuclei merge. Dim cells disappear. Debris becomes a population. A model tuned on one plate inherits the illumination and staining of that plate.

We build and validate image-analysis workflows around the measurement the experiment actually needs.

• Cellpose or StarDist works on one image set and fails on another batch.
• Manual thresholds change between operators or sessions.
• Dense objects merge and small objects vanish.
• Background correction changes the apparent treatment effect.
• The team has thousands of fields and no systematic QC.
• A global Dice score looks acceptable while counts or phenotypes are biased.
• The current workflow exports masks but cannot trace measurements back to source images.

Segmentation and detection

Cells, nuclei, colonies, puncta, fluorescent objects and other structures with a defined visual target. Candidate methods may include Cellpose, StarDist, classical OpenCV, watershed, connected components or a task-specific model.

Quantification

Counts, area, intensity, morphology, colocalisation, neighbourhoods and spatial distributions. The measurement definition is written before batch processing.

Validation

Object-level matches, misses, false detections, merges, splits, boundary errors and performance by image condition. Metric choice follows the downstream question.

Batch and image QC

Focus, exposure, illumination, saturation, field coverage, debris, staining shift and acquisition metadata.

Reproducible processing

Batch runners, configuration, image manifests, QC reports, mask exports, measurement tables and links back to source images.

StarDist represents objects as star-convex polygons or polyhedra. That is a strong fit for many nuclei and a poor fit for branching structures. Cellpose predicts spatial flows and has broad generalist coverage, but a generalist model still needs testing on the client’s stain, magnification, density and failure cases.

Classical methods remain useful. A stable fluorescence spot on a controlled background may be handled better by flat-field correction, a local threshold and connected-component rules than by a neural network with opaque failure modes.

A small annotation set should be selected across the conditions most likely to break the method:

• dim and bright fields;
• sparse and dense regions;
• edge-of-well and centre fields;
• treatment and control;
• early and late acquisition dates;
• different operators, instruments or staining batches.

Annotating 20 easy images at random gives a precise answer to the wrong question.

• Image and metadata audit
• Baseline comparison across candidate methods
• Annotation sampling plan
• Segmentation or detection pipeline
• Object-level error analysis
• Batch and image QC
• Measurement table linked to source images and masks
• Reproducible code, configuration and handover

Which image-analysis tools do you use?

Cellpose, StarDist, OpenCV, scikit-image, PyTorch and ImageJ-compatible outputs are common choices. The tool follows the object geometry, image formation and handover environment.

Do we need full ground-truth masks?

Often no. The amount and type of annotation depend on the measurement. Counting requires reliable object detection and separation. Fine morphology requires stronger boundary annotation. A stratified small set can expose failure faster than a large convenient set.

Can you process clinical images?

Research analysis of appropriately governed data can be scoped. Diagnostic use, clinical decision support and regulated deployment require separate review of intended purpose, governance and insurance.

Can the data stay on our systems?

Yes, where the client environment supports the required tools and access. Derived examples or figures leave the environment only under the agreed data terms.

Send three ordinary images and three failures.

Include pixel size, channel meaning, acquisition conditions, the measurement required and any existing masks or manual counts. Difficult examples are more useful than a folder of polished fields.

Scope an image-analysis review

• Validate cell segmentation with limited ground truth

How Services engagements are scoped