Selected work

Challenge

Biomarker scoring models need ground truth that is consistent across slides, sites and annotators.

Approach

Regional annotation marked tumour and tissue compartments, single-cell annotation labelled individual tumour cells, and for some programmes we reviewed and corrected the regions a model had drawn. Every annotation followed the client’s written guidelines, and work by the wider PN team was reviewed by an ABPath- and/or FRCPath-certified senior pathologist before delivery. The programmes were:

• TROP2 in urothelial carcinoma
• CLDN18.2 in biliary tract and pancreatic carcinoma
• GPC3 in hepatocellular carcinoma
• FRα in non-small cell lung carcinoma
• B7-H3, STEAP2 and PSMA in prostate carcinoma
• CLDN3, GUCY2C and cMET in colorectal carcinoma
• Residual viable tumour after neoadjuvant therapy in non-small cell lung carcinoma
• Cell viability in cell pellet preparations

Outcome

The programme delivered quality-controlled, expert-annotated datasets for training, validating and testing oncology biomarker AI models.

Challenge

A collection of slides is not yet a dataset. Before annotation starts, each case needs a confirmed diagnosis, the right tissue, metadata that matches the image and a clear decision on whether it meets the study’s inclusion criteria.

Approach

Subspecialist pathologists reviewed the case metadata and whole-slide images for each candidate case, verified the diagnosis and recorded a structured assessment of tumour morphology, variant, grade and staining, with a judgement on each specimen’s suitability against the inclusion criteria.

Outcome

The programme delivered documented, expert-verified pathology assessments across eleven cancer types, giving a consistent basis for dataset qualification and later AI development.

Challenge

Reviewing a pathology report, a research dataset or a laboratory audit asks an AI agent to find the right information, check it and reach a sound conclusion, often across several files and applications. The programme needed realistic tasks of this kind, each with defined success criteria, supporting files and a structured assessment of how each model performed.

Approach

Each team member designed tasks from their own professional field, and most came from pathology and cancer research. Agents extracted tumour stage and receptor status from reports, quality-checked synoptic and melanoma reports, reviewed prostate grade groups and frozen section discordances, prepared cases for multidisciplinary meetings, validated cancer dataset metadata, and searched and summarised the biomarker literature. Other tasks covered laboratory operations and business workflows. Every task was run on several AI models and scored against predefined checkpoints, with a written account of what each model did well and where it failed. PN recruited and coordinated the team, set consistent standards for task design and scoring, reviewed each evaluation before submission and built internal tools to check quality and completeness.

Outcome

PN delivered 129 quality-controlled evaluations covering over 500 model runs, giving structured evidence of how well each agent completed realistic clinical and research workflows and where it failed.

Challenge

Language models are increasingly asked to interpret pathology reports, and judging whether an answer is accurate and safe needs a pathologist’s standard.

Approach

Pathology report benchmark. For 20 surgical pathology reports across gastrointestinal, breast, genitourinary, gynaecological, skin, head and neck, thoracic, salivary and liver pathology, we wrote the instructions given to the model, a reference answer and a scoring rubric of at least ten criteria, each weighted by clinical importance. Each case was run against five or more models.

Expert medical tasks. Between December 2025 and April 2026, in a separate programme, we wrote 35 expert medical tasks, each with a realistic clinical request, a reference deliverable and its sources, and reviewed another 20 written by other experts against the client’s reviewer checklist.

Outcome

The benchmark gives AI developers a reusable way to compare different models’ answers against a consultant pathologist’s standard, and the expert tasks gave them realistic, referenced clinical work for training.