Home/Services/AI image analysis
Capability
AI image analysis
Threshold-based analysis works until the tissue stops cooperating. Dense nuclei, weak or uneven staining, necrosis, folds, morphologically heterogeneous tumour — that is where trained models earn their place. We build them on your material and we tell you honestly where they fail.
What the models actually do
- Nuclear segmentation
Separating touching and overlapping nuclei in dense regions, where a simple intensity threshold merges neighbouring cells and undercounts the population. Accuracy here sets the ceiling for every downstream measurement.
- Cell segmentation and phenotyping
Assigning membrane and cytoplasmic signal to the correct nucleus, then classifying each cell by its marker combination — the basis of any multiplex phenotype call.
- Tissue classification
Identifying tissue morphologies by colour, texture and contextual features: tumour, stroma, necrosis, immune aggregates, glandular structures. This removes the need for exhaustive manual annotation on every slide.
- Spatial analysis
Distances between cell types, infiltration across a defined tumour border, neighbourhood composition, and co-localisation of activation markers — the questions that only become answerable once phenotyping is reliable.
- Artefact and exclusion models
Automatically flagging folds, bubbles, out-of-focus areas, pen marks and edge effects so they never enter the analysed area.
How we know a model is good enough
Every model is developed on a training subset and evaluated on held-out images it has never seen, against annotations made by a pathologist. We report agreement, not just an accuracy figure, and we show you the cases where the model and the pathologist disagreed.
Ground truth
Pathologist-annotated held-out set
Reported metrics
Precision, recall, F1, concordance
Failure review
Disagreement cases returned to you
Version control
Every model versioned and archived
Models are tools, not oracles. Where a model performs poorly on a tissue type, a stain or a scanner we say so and either retrain, restrict the analysis scope, or recommend a different approach. We do not ship a number we would not defend in a review meeting.
Built to run at study scale
Infra
Scalable cloud
Whole-slide analysis runs on elastic compute, so a 2,000-slide cohort does not take twenty times longer than a 100-slide one.
Infra
Access controlled
Role-based access, encrypted storage and transfer, UK/EU data residency, and retention terms agreed in the contract.
Infra
Reproducible runs
Container-pinned software versions and recorded parameters, so a run can be repeated years later and give the same answer.
Got tissue that breaks conventional analysis?
Those are the interesting ones. Send a few of the worst slides and we will tell you what is recoverable.