Digital pathology AI sits on a regulatory fault line. When the software analyses whole-slide images derived from in vitro specimens to provide information for diagnosis, its primary framework is often the IVDR, not the MDR. But the qualification depends on intended purpose under Article 2(12), and the classification under MDR Annex VIII Rule 11 still matters where the software is framed as a medical device rather than an IVD. Getting this wrong at founding costs a year.

By Tibor Zechmeister and Felix Lenhard.

TL;DR

  • Digital pathology software analysing images of in vitro specimens (tissue, cells, body fluids) for diagnostic information is typically governed by IVDR (Regulation (EU) 2017/746), not MDR.
  • Where a pathology tool is framed as a medical device under Article 2(1) rather than an IVD — for example, a workflow or visualisation tool without diagnostic claims on specimens — MDR applies.
  • Rule 11 under MDR Annex VIII pushes diagnostic decision support to at least Class IIa, with Class IIb or III realistic for primary diagnosis claims in cancer pathology.
  • The boundary between "primary diagnosis" (highest risk) and "decision support" (lower risk) is decided by the wording of intended purpose — not by marketing aspiration.
  • Scanner validation is a manufacturer obligation: your intended purpose must specify which whole-slide scanners are in scope, and evidence must cover each.
  • Founders must pick the correct regulation (MDR or IVDR) before writing the technical file. Changing your mind mid-project rewrites the dossier.

Why this matters

A digital pathology startup spends eighteen months building a breast cancer tissue classification model. They pitch it as "AI that reads your biopsy slides." They assume MDR applies because they know more about MDR than IVDR. Their notified body contact — correctly — tells them the device looks like an IVD, and their dossier is pointed at the wrong regulation. They now have a choice: reframe the intended purpose to stay in MDR scope, which narrows the claims dramatically, or rebuild the regulatory strategy under IVDR, with a different notified body designation, different classification rules (IVDR Rule 3), and different clinical performance expectations.

Either path is viable. Neither is free. And the decision should have been made on day one, not month eighteen.

Digital pathology is the domain where founders most frequently confuse MDR and IVDR, and where the Subtract to Ship discipline — pick your regulatory scope with intent, match your claims to it, defend the boundary — matters most.

What MDR actually says

Article 2(1). A medical device is an instrument, software, or other article intended by the manufacturer for specific medical purposes. The definition is broad, but it explicitly excludes in vitro diagnostic medical devices, which fall under IVDR.

Article 2(2). An in vitro diagnostic medical device is a device intended to be used in vitro for the examination of specimens, including blood and tissue donations, derived from the human body, solely or principally for the purpose of providing information — for diagnosis, monitoring, and similar purposes. (Full text in Regulation (EU) 2017/746 Article 2(2); the MDR's scope exclusion is in MDR Article 1.)

This is the fork. Software that analyses images of tissue specimens taken from the body, solely or principally to provide diagnostic information, is an IVD. That is the default framing for most AI pathology tools.

Article 2(12). Intended purpose is whatever the manufacturer declares on the label, in IFU, in promotional material, or in the clinical evaluation. This sentence is as decisive for pathology as for any other AI MedTech domain.

Annex VIII Rule 11. If — and only if — a pathology software tool is framed as a medical device under MDR (not an IVD), Rule 11 applies. Decision support with diagnostic or therapeutic purpose is Class IIa, or Class IIb where wrong output could cause serious deterioration or surgical intervention, or Class III where wrong output could cause death or irreversible deterioration. For cancer pathology, the severity of wrong decisions puts most meaningful claims at Class IIb or above.

MDCG 2019-11 Rev.1 (June 2025). The guidance applies to software qualification and Rule 11 classification under MDR. It is the reference for MDR-path digital pathology. IVDR-path pathology is governed by IVDR Rule 3 and separate MDCG guidance.

The practical upshot. A founder's first decision is not "what class are we?" It is "are we MDR or IVDR?" That decision is determined by intended purpose and the nature of the specimens examined. Once it is made, the relevant classification rules and evidence frameworks follow. Flipping between them is not a tweak — it is a different dossier.

A worked example

A digital pathology startup has three product framings in front of the founders:

Framing A — Primary diagnosis on H&E breast biopsies. Intended purpose: "Provides a primary diagnostic classification of invasive ductal carcinoma from whole-slide images of H&E-stained breast biopsy sections." Marketing: "AI that diagnoses breast cancer from tissue slides."

This is an IVD, not an MDR device. The software's sole purpose is providing diagnostic information from in vitro examination of tissue specimens. IVDR Rule 3 applies (not MDR Rule 11), and for cancer diagnosis, IVDR classification typically puts this at Class C or Class D with notified body involvement. This post will not solve your full IVDR path, but it will stop you filing the wrong dossier.

Framing B — Decision support to pathologist on H&E breast biopsies. Intended purpose: "Supports the pathologist's interpretation of whole-slide images of H&E-stained breast biopsy sections by highlighting regions suggestive of invasive carcinoma and providing a confidence estimate. The pathologist remains responsible for the final diagnosis."

This is still likely IVDR, because the information provided is for diagnostic purposes from in vitro specimens. The "decision support" framing does not remove it from IVDR scope. It might moderate the IVDR classification in some framings, but it does not move the device into MDR. Many founders try this workaround and fail their first notified body conversation.

Framing C — Pathologist workflow tool, no diagnostic claims. Intended purpose: "Provides visualisation, measurement, and annotation tools for whole-slide images to support pathologist workflow. The device does not provide diagnostic information or characterise findings."

This is more likely an MDR medical device (or in some cases not a medical device at all, depending on whether any medical purpose is claimed). If medical purpose is claimed, Rule 11 applies. Without diagnostic or therapeutic decision information, Class I may be defensible; with any decision-supporting output, at least Class IIa.

Same underlying technology. Three entirely different regulatory worlds — driven by intended purpose under Article 2(12).

Scanner scope. Whichever framing you choose, your intended purpose must specify which whole-slide scanners, staining protocols, and imaging parameters are in scope. A model validated on one scanner vendor's output is not automatically valid on another's. Scanner validation is part of your technical file, and notified bodies ask about it.

The Subtract to Ship playbook

Step 1 — Decide MDR or IVDR before you write any dossier. This is the single most expensive decision in a digital pathology startup. Make it deliberately, in writing, with justification referencing MDR Article 2(1), IVDR Article 2(2), and the actual specimens and intended purpose. Sign it, date it, revisit only if intended purpose changes.

Step 2 — Write the intended purpose against the regulation you chose. If you are IVDR, your intended purpose language must read like an IVD intended purpose. If you are MDR, it must read like a medical device intended purpose. Mixed language is a red flag to every notified body reviewer.

Step 3 — Match marketing to the chosen scope. A pathology startup with "AI breast cancer diagnosis" on the landing page and an MDR workflow-tool technical file is not going to survive due diligence, let alone a notified body audit. Pick your lane and keep marketing in it.

Step 4 — Scope scanners narrowly, then expand. List the exact whole-slide scanners, magnifications, staining protocols, and tissue types in your intended purpose. Validate against each. Add scanners as post-market changes with documented equivalence or fresh validation. Broad scanner claims from day one are a clinical evidence disaster.

Step 5 — Build the risk file around real pathology error modes. Under EN ISO 14971, the hazards to model include: false negative on invasive cancer, false positive leading to unnecessary treatment, domain shift between scanner vendors, staining variation, rare morphology underperformance, and demographic subgroup effects. Each needs estimation and controls.

Step 6 — Plan clinical performance evaluation against the actual workflow. Retrospective standalone accuracy is not enough. For MDR-path digital pathology tools, plan reader studies showing the tool's impact on actual pathologist decisions. For IVDR-path tools, plan clinical performance studies as required under IVDR Annex XIII.

Step 7 — Subtract claims, narrow population, stage the launch. Start with the narrowest defensible intended purpose — one scanner, one tissue type, one decision context. Expand after CE/IVDR certificate. Every claim you subtract at the start is a month of clinical work you do not have to redo later.

Reality Check

  1. Have you made, in writing, the MDR vs. IVDR decision for your device, with justification traceable to Article 2(1), IVDR Article 2(2), and your actual specimens?
  2. Does your landing page match the regulation you chose, or does it describe a product that sits in the other regulation?
  3. Is your intended purpose specific about scanners, staining, tissue types, and decision context — or is it aspirational and broad?
  4. If you are IVDR, have you identified your notified body's IVDR designation and started the engagement? (Many MDR notified bodies are not designated for IVDR.)
  5. Does your risk file identify domain shift across scanner vendors and staining variation as hazards, with controls?
  6. Is your clinical performance plan built on workflow-realistic studies, not standalone retrospective accuracy?
  7. Are you willing to stage your launch — one tissue type, one scanner first — to make the evidence work tractable?
  8. If an auditor asked tomorrow why you chose MDR over IVDR (or vice versa), could you produce a one-page signed justification?

Frequently Asked Questions

If we analyse whole-slide images rather than the physical tissue, aren't we just image software under MDR? No. The regulation looks at the purpose of the examination and the nature of the specimen. Whole-slide images of tissue specimens examined for diagnostic information fall under IVDR scope, because they derive from in vitro examination of specimens. The image is a representation of the specimen, not a different thing.

Can we avoid IVDR by being "decision support only"? Rarely. IVDR scope is defined by purpose, not by the degree of automation. A decision-support tool that provides diagnostic information from in vitro specimens is typically still an IVD.

What if our pathology tool is used by pathologists on their own archived images without touching the specimen? The regulatory analysis still turns on intended purpose and the nature of the information provided. Archived whole-slide images derived from in vitro specimens do not leave IVDR scope just because the specimen is no longer involved.

How do we handle multiple scanner vendors? Specify each in the intended purpose and validate each. Post-market addition of new scanners is a change control event, with fresh validation or documented equivalence. Do not assume scanner-agnostic performance.

Are pathology AI tools ever outside both MDR and IVDR? Yes, if they claim no medical purpose at all — pure research tools, education tools, or visualisation software with no diagnostic intent. But marketing for sale into clinical workflows almost always crosses a line.

Should we talk to a notified body before we lock our intended purpose? Yes. For digital pathology specifically, an early conversation with a notified body designated under the relevant regulation can save a year. Bring a draft intended purpose and a written MDR/IVDR decision memo.

Sources

  1. Regulation (EU) 2017/745 on medical devices, consolidated text. Article 1, Article 2(1), Article 2(12), Annex VIII Rule 11.
  2. Regulation (EU) 2017/746 on in vitro diagnostic medical devices. Article 2(2), classification rules.
  3. MDCG 2019-11 Rev.1 (October 2019, Rev.1 June 2025) — Guidance on qualification and classification of software in Regulation (EU) 2017/745.
  4. EN 62304:2006+A1:2015 — Medical device software — Software life cycle processes.
  5. EN ISO 14971:2019+A11:2021 — Application of risk management to medical devices.
  6. EN ISO 13485:2016+A11:2021 — Quality management systems for medical devices.