Reading the Regulatory Map: How FDA Clears AI in Pathology, and Why So Few Make It Through

If you've spent any time on the exhibit floor at a digital pathology meeting lately, you'd be forgiven for thinking the field is drowning in cleared AI. Every booth has an algorithm; every algorithm has a demo. But the regulatory ledger tells a more sober story. As of April 2026, the FDA has authorized roughly 51 AI-enabled digital pathology devices, the overwhelming majority through the 510(k) or De Novo pathways [1]. Fifty-one. For a field that markets itself as being on the cusp of transformation, that's a small number — and understanding why it's so small is the most useful orientation any pathologist can have right now.

The development: a maturing pathway, not a floodgate

The story here isn't a single breakthrough clearance. It's the shape of the pathway itself becoming legible. The FDA now maintains a live, public database of AI-enabled medical devices under its Software as a Medical Device (SaMD) umbrella, and that list is worth treating as the primary reference document for anyone evaluating a tool [2]. It distinguishes devices by regulatory route — De Novo, 510(k), and the rarer premarket approval (PMA) — and it's the closest thing we have to ground truth on what's actually authorized versus what's merely in a press release.

The distinction between those pathways matters more than it might first appear. A De Novo authorization is what you get when a device is genuinely novel and low-to-moderate risk with no existing predicate — it creates a new device classification. A 510(k) clearance, by contrast, rests on demonstrating substantial equivalence to a device already on the market. Once a De Novo goes through, it can serve as the predicate for subsequent 510(k) submissions, which is how a category slowly opens up. The first tool in a class does the heavy regulatory lifting; the ones that follow ride a somewhat shorter road [1].

The most consequential recent example of that dynamic is PathAI's AISight Dx platform, cleared by the FDA in 2025 for primary diagnosis in digital pathology [3]. The word primary is doing a lot of work in that sentence. Most cleared pathology AI to date has been positioned as decision support — a second read, a triage aid, a quantification assistant. A platform cleared for primary diagnosis is operating much closer to the diagnostic act itself, and it establishes a predicate precedent that's relevant to future submissions touching breast, prostate, colon, and lung tissue [3]. That's why a single platform clearance is worth more than its immediate use case suggests: it reshapes the map for everyone downstream.

The diagnostic problem it addresses

Strip away the regulatory vocabulary and the underlying problem is one every practicing pathologist knows intimately. The tissues at the center of this SaMD ecosystem — breast, lung, colon, prostate, ovarian, and lymphoma specimens — are exactly the high-volume, high-consequence areas where reproducibility and workload pressure collide [1][2]. Prostate core biopsies arrive in sets of a dozen or more, most of them benign, and the diagnostic signal you're hunting for may occupy a fraction of a millimeter. Breast and colon cases carry enormous downstream therapeutic weight. Ovarian and lymphoma work demands pattern recognition across morphologic categories that even experienced observers grade with meaningful interobserver variability.

AI-enabled SaMD is aimed squarely at that combination of volume and variability. The promise is consistency — a tireless first pass that flags, quantifies, or in the case of a primary-diagnosis platform, participates directly in the read. But — and this is the point the regulatory framework enforces — a tool that participates in diagnosis of these cancers is making claims with real clinical stakes. The bar should be high, and the pathway is designed to make it high.

The evidence

Here's where measured language is not just editorial style but epistemic necessity. The confirmed developments establish what has been authorized and through which pathway, but they don't hand us a trove of head-to-head sensitivity and specificity figures to recite. What the record does show is a regulatory system functioning as a filter. Fifty-one authorizations through April 2026, against a development pipeline that's clearly far larger, tells you that most tools either haven't been submitted, haven't cleared, or are still working through validation [1].

That gap is the evidence, in a sense. The FDA's device list is structured to support lifecycle management — meaning clearance isn't a one-time event but the start of an ongoing obligation to monitor performance [2]. For AI, whose behavior can drift as scanners, staining protocols, and patient populations shift, that lifecycle framing is arguably the most important feature of the whole apparatus. A clearance validated on one institution's slides is not a guarantee of performance on yours. The evidence supporting any given device lives in its specific submission, and the responsible move for a laboratory is to read that submission rather than the marketing.

AISight Dx's clearance for primary diagnosis represents the strongest validation signal in the current set, precisely because the FDA doesn't authorize a primary-diagnosis role casually [3]. But even there, the appropriate reading is narrow: the clearance attests to the claims the FDA reviewed, not to universal performance across every cancer, every scanner, and every workflow.

Where it stands

So where does this leave a pathologist trying to make sense of it all? On solid but modest ground. The pathways are real and increasingly well-trodden: De Novo to open a category, 510(k) to populate it, PMA for the highest-risk applications [1][2]. A meaningful precedent now exists for AI participating in primary diagnosis rather than sitting safely in the decision-support lane [3]. And the FDA's public database gives every laboratory a way to check claims against authorizations directly [2].

The open questions are substantial and worth naming plainly. This is a fast-moving area, and the small clearance count reflects both a genuinely high bar and the sheer difficulty of validating AI across the messy heterogeneity of real-world pathology. Lifecycle monitoring — how cleared tools are re-evaluated as they encounter new data — remains the frontier where policy and practice are still being worked out. And the therapy connection, where AI-derived findings on these breast, lung, colon, prostate, ovarian, and lymphoma specimens may inform eligibility for a drug class, is exactly where clearance scope needs to be read most carefully. A tool cleared to quantify a feature is not thereby cleared to direct treatment.

The map is getting clearer. It is not yet the territory. For now, the most valuable skill isn't picking the flashiest algorithm — it's reading the clearance and knowing precisely what the FDA did, and didn't, authorize.

---

References

1. Shrestha Y. AI/ML in Digital Pathology and the Software-as-an-IVD Paradigm: 2026 Snapshot. Innolitics (primary industry regulatory analysis), 2026. https://innolitics.com/articles/ai-ml-in-digital-pathology-and-the-software-as-an-ivd-paradigm-snapshot

2. U.S. Food and Drug Administration. Artificial Intelligence-Enabled Medical Devices. FDA.gov — Software as a Medical Device (SaMD) section, 2026. https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-enabled-medical-devices

3. PathAI. PathAI Receives FDA Clearance for AISight® Dx Platform for Primary Diagnosis. PathAI Press Release / FDA clearance announcement, 2025. https://www.pathai.com/news/pathai-receives-fda-clearance-for-aisight-dx-platform-for-primary-diagnosis