Digital pathology & slide analysis

Pathologists diagnose cancer and many other diseases by reading glass slides under a microscope, a scarce, labour-intensive skill. Once slides are scanned into gigapixel whole-slide images (WSI), AI can pre-screen for cancer, count cells, grade tumours, score biomarkers and prioritise urgent cases. The pathologist stays responsible for the diagnosis; AI acts as a second reader and a quantification tool. In the EU this means CE-marked in-vitro diagnostic software under IVDR plus AI Act high-risk duties.

Problem & value

• Today: tissue is cut, stained and examined by a pathologist on a microscope. Pathologist shortages, rising cancer caseloads, and growing biomarker panels (immunotherapy, targeted therapy) create backlogs. Grading and manual counting (mitoses, Ki-67, PD-L1) show inter-observer variability.
• AI value: second-read safety net (e.g. missed small cancer foci in prostate biopsies), triage of negative vs. suspicious cases, reproducible quantification, fewer re-cuts and ancillary stains, and tumour-board-ready measurements. Published reader studies and vendor-reported deployments show workload and consistency benefits; effect sizes depend strongly on tissue type and site, so measure locally.
• Prerequisite: the benefit only materialises after a digital workflow exists (scanners, image management system, storage, LIS integration). Digitisation itself is the larger investment; AI is an add-on to it [1][2].

Variants

• Diagnostic decision support: prostate, breast, colorectal, lung, gastric, skin biopsies (cancer detection, grading).
• Quantitative IHC/biomarker scoring: HER2, Ki-67, ER/PR, PD-L1, claudin-18.2; companion-diagnostic style.
• Triage and QC: flag blur, folds, missing tissue, wrong stain; prioritise worklist.
• Cytology: cervical screening and urine/thyroid cytology (see also HLTH-026 for lab-diagnostics).
• Pharma / research use (RUO): biomarker discovery, toxicologic pathology, clinical-trial endpoints (see HLTH-017, SCI-009). Not subject to the clinical-diagnostic rules but must not be reused clinically.
• Computational pathology foundation models: slide-level embeddings for retrieval, molecular-status prediction, and report drafting.
• Veterinary pathology: same technology outside IVDR (see HLTH-016).

How it works

1. glass slide
2. scanner (WSI, DICOM/vendor format)
3. image management system / PACS
4. quality control (focus, tissue detection, stain check)
5. tiling into patches (e.g. 224-512 px at 20x/40x)
6. foundation-model encoder
7. patch embeddings
8. task head: multiple-instance learning (attention aggregation) for slide-level labels | segmentation/detection for cells and regions
9. heatmap + score + confidence
10. viewer overlay in pathology workstation
11. pathologist review, accepts/edits/rejects
12. signed-out report in LIS

• Foundation models: self-supervised vision transformers pre-trained on hundreds of thousands to millions of slides produce reusable patch features; downstream heads need far fewer labels than training from scratch.
• Aggregation: because a slide has no pixel-level labels, weakly supervised multiple-instance learning is standard.
• Human-in-the-loop: the pathologist makes and signs the diagnosis. The AI result is shown as overlay or prompt; disagreement is logged. Rare entities and unusual findings always need a human.

Data preparation

The data are whole-slide images (WSI) from slide scanners, typically 1-5 GB each, plus case data from the LIS and pathology reports. Many institutes still work with glass slides, so digitisation itself is the first step. Typical gaps: no or partial slide scanning, proprietary scanner formats, heterogeneous staining and scanner colour profiles, out-of-focus areas, missing barcodes or case links, free-text reports without structure, and very large storage needs.

Preparation steps: (1) establish routine scanning and link every slide to its LIS case through barcodes; (2) convert proprietary formats to a common one and store them in an image management system or a PACS with pathology support; (3) check scan quality automatically (focus, tissue detection); (4) agree stain and scanner standards or validate across them; (5) for training or validation, build annotated reference sets with pathologists (PREP-018, PREP-019) and harvest labels from signed reports with care (PREP-020, PREP-021); (6) pseudonymise slide labels and metadata (PREP-022); (7) plan storage and retention; for paper-based case archives see PREP-006.

Effort: high. Scanning workflow, storage and annotation dominate; buying a certified product reduces annotation but not the digitisation effort. See GUIDE-003 and DATA-003.

Implementation paths

Which path for whom

• Enterprise (pathology institute, lab group): Path A, because certified digital-pathology AI integrated with the slide scanner and LIS exists and carries the regulatory burden.
• Public (university or municipal pathology): Path A with certified products, because procurement rules favour auditable suppliers, with research use built on open tools.
• Developer: Path C, because open-source whole-slide tooling and foundation models run on own GPUs and keep patient slides in-house.

Path A — Buy (recommended)

• Clinical, CE-marked AI (check current IVDR/MDR status per product): Paige (acquired by Tempus AI in 2025; Paige Prostate and platform), Ibex Medical Analytics (Galen Prostate/Breast), Aiforia, Visiopharm, Indica Labs (HALO AP), Philips (IntelliSite/Pathology Suite with AI apps), Roche (navify Digital Pathology and algorithm suite), Mindpeak (DE, breast IHC scoring), Aignostics (DE, spin-off of Charité), Owkin (FR), Deep Bio, Lunit (SCOPE). EU-based vendors: Mindpeak, Aignostics, Aiforia (FI), Owkin, Visiopharm (DK).
• Image management / platform: Sectra (SE), Proscia Concentriq, Philips, Leica/Aperio (Leica Biosystems), Hamamatsu, 3DHistech (HU) scanners and viewers; open standards are DICOM-WSI and vendor-neutral viewers.
• Choose A when: the goal is clinical use. Only a certified product gives the manufacturer's intended-purpose, performance evidence and liability that a laboratory needs.

Path B — Build on cloud APIs

• Cloud storage/DICOM stores with EU regions (e.g. Azure Health Data Services, Google Cloud Healthcare API, AWS HealthImaging) and managed GPU inference for research pipelines.
• General multimodal LLMs can draft report text from structured findings, but are not suitable to interpret slides for diagnosis. Use only for research or report-language assistance (see HLTH-011).
• Choose B for research, pharma RUO workflows and infrastructure; not for undocumented clinical use.

Path C — Self-host (open source)

• Viewers/analysis: QuPath (open source, widely used for cell detection and annotation), OpenSlide, openslide-python, TIAToolbox, Cytomine, Orthanc/dcm4chee with DICOM-WSI, MONAI.
• Open foundation models (weights often under research or gated licences, check terms): UNI/UNI2 and CONCH (Mahmood Lab, Harvard), Virchow/Virchow2 (Paige/Microsoft), Prov-GigaPath (Microsoft/Providence), CTransPath, Phikon (Owkin, FR), H-optimus (Bioptimus, FR) and Atlas (Aignostics). Pipelines such as CLAM, TRIDENT and Kaiko's eva support slide-level heads.
• Hardware: HW2 can run patch-embedding on moderate volumes; HW3 (48-141 GB GPU) for fast embedding of whole archives and larger encoders. Storage is the main constraint: a single WSI is typically hundreds of MB to several GB, so an archive needs tiered storage.
• Self-built tools used for diagnosis count as in-house devices (IVDR Art. 5(5), see risks).

Path D — Custom training

Worth it for rare entities, local stain/scanner conditions or a research question. Requires hundreds to thousands of expert-annotated or outcome-linked slides across multiple scanners and sites, slide-level labels plus a held-out external validation set. Fine-tune a foundation-model head rather than train an encoder. Stain normalisation and augmentation reduce scanner shift. Data governance (ethics approval, consent or legal basis, pseudonymisation) is usually the hardest step.

State of the art (2026)

• Foundation models have become the default backbone for computational pathology; large public benchmarks and independent evaluations show strong cancer detection and subtyping, with weaker and less consistent results on rare cancers, prognosis and molecular-status prediction.
• Vision-language and agentic models: CONCH-style slide-text models enable retrieval and zero-shot classification; multimodal pathology copilots are emerging but are mostly research-grade.
• Regulatory milestones: the first AI pathology products received FDA De Novo / clearance (Paige Prostate in 2021) and several CE-marked products exist under IVDD/IVDR; digital primary diagnosis is now accepted practice in many countries and large-scale national programmes exist [1][3].
• Limits and failure modes: domain shift (new scanner, stain protocol, laboratory), artefacts (folds, pen marks), over-trust/automation bias, rare tumours, poor calibration, shortcut learning from site-specific slide features, and high storage and compute cost. Performance in one laboratory does not transfer automatically; local verification is essential.

Shared resources & combinations

• Imaging stack shared with HLTH-001 (radiology), HLTH-006 (derm/ophthalmology), HLTH-026 (lab diagnostics), HLTH-014 (dental): DICOM stores, viewers, GPU inference servers (HW3) and MLOps for medical-device software.
• Foundation-model embeddings combine with genomics (SCI-009) and clinical records (HLTH-018) for multimodal prognosis; with HLTH-009 for trial patient matching; with HLTH-017 for submission documentation.
• Terminology and interoperability: SNOMED CT/LOINC, synoptic reporting, FHIR/DICOM mapping (HLTH-028).
• Synthetic data / augmentation: see DATA-005 (use cautiously: synthetic slides do not replace validation).

Effort, cost & KPIs

• Path A: 6-18 months including scanner roll-out, IT integration, validation and training. Largest costs: scanners, storage, change management; per-case or licence fees for AI modules.
• Path B/C: research pipeline in 2-4 months; clinical-grade in-house development takes a year or more with QA and regulatory work.
• KPIs: sensitivity/specificity vs. consensus diagnosis, agreement (kappa) between pathologists with and without AI, turnaround time, share of cases needing re-cuts/extra stains, slides scanned per day, scan failure and rescan rate, pathologist time per case, concordance monitoring after go-live.

People & change

Management and the head of pathology own the decision, because the case usually requires a scanner investment and a changed workflow. Involve pathologists and biomedical lab technologists early, IT for storage, scanner and LIS integration, and IT security for the network. The DPO assesses legal basis, AVV with the AI supplier and a DPIA, which is likely for health data at scale (GUIDE-005). Data owners of the slide archive and, for research reuse, the ethics committee join before data leave the institute.

Works council: depends. Information in good time is required where one exists (§90(1) no. 3 BetrVG). Co-determination under §87(1) no. 6 BetrVG becomes likely if the system records which pathologist reviewed or changed what, because objective suitability for monitoring is enough. Public institutes involve the Personalrat (GUIDE-004).

AI literacy: support pathologists and technologists with training on typical errors, out-of-distribution slides and automation bias (GUIDE-006). Pick a pilot group of two to three pathologists and one organ type, run the AI as a second reader alongside normal diagnosis (GUIDE-009). Concerns are liability, deskilling and fear of replacement: keep the pathologist as the sole diagnostician, make every AI marking verifiable on the slide, and communicate openly that the tool does not rate individuals (GUIDE-008).

Risks, regulation & compliance

• EU AI Act: diagnostic AI is a medical device/IVD requiring third-party conformity assessment, hence high-risk under Art. 6(1) with Annex I (MDR/IVDR). High-risk obligations for such products apply from 2 August 2028 (deferred by the Digital Omnibus on AI, Reg. (EU) 2026/1744). Deployer duties: human oversight, log keeping, AI literacy (Art. 4) [4].
• IVDR (EU 2017/746): software with a diagnostic purpose is an IVD; most such software is class B or C, so a notified body is involved. Laboratories building their own tools fall under the in-house exemption (Art. 5(5)) with conditions (QMS, justification, no equivalent CE product on the market) [5].
• GDPR: slides and linked data are health data (Art. 9); needs legal basis, DPIA, pseudonymisation, processor contracts, EU hosting/no unsafe transfers. Secondary use for training needs consent, research exemption or national law (e.g. EHDS, national health data laws); WSIs can contain labels with identifiers.
• Other: European Health Data Space regulation for secondary use; national pathology standards and accreditation (ISO 15189); clinical validation per guidance such as the CAP/College of American Pathologists and RCPath digital pathology guidance.
• Mitigations: local validation before go-live, ongoing performance monitoring, scanner/stain change control, pathologist sign-off, audit log of AI result vs. final diagnosis, bias checks across demographics.

Getting started

1. Assess readiness: case volume, subspecialties, existing scanners, LIS and storage; define one use case (e.g. prostate biopsy second read).
2. Pilot a CE-marked product on retrospective cases from your own lab; measure sensitivity, specificity and time.
3. Set up governance: DPIA, ethics/legal basis, intended-purpose and oversight rules, validation protocol.
4. For research, try QuPath plus a public foundation model (UNI/CONCH/Virchow2) on a small public dataset such as TCGA or Camelyon.
5. Go live in shadow mode first, then supervised use, and monitor concordance monthly.

Sources

1. College of American Pathologists / RCPath digital pathology validation guidance — https://www.rcpath.org/profession/digital-pathology.html
2. Regulation (EU) 2017/746 (IVDR) — https://eur-lex.europa.eu/eli/reg/2017/746/oj
3. FDA — De Novo for Paige Prostate (DEN200080) — https://www.accessdata.fda.gov/scripts/cdrh/cfdocs/cfpmn/denovo.cfm
4. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
5. QuPath open-source software — https://qupath.github.io/
6. UNI / CONCH foundation models (Mahmood Lab) — https://github.com/mahmoodlab/UNI
7. Virchow2 foundation model (Paige) — https://huggingface.co/paige-ai/Virchow2
8. Prov-GigaPath, Nature 2024 — https://www.nature.com/articles/s41586-024-07441-w
9. European Health Data Space regulation — https://health.ec.europa.eu/ehealth-digital-health-and-care/european-health-data-space_en