Blog · Analysis · Last reviewed June 23, 2026

The Pathology Model Becomes the Second Reader

AI-assisted pathology does not replace the microscope all at once. It becomes a second reader that can redirect attention, uncertainty, workflow, and diagnostic responsibility.

For this essay, a pathology second reader is the whole clinical workflow that lets software inspect digitized tissue and return a mark, coordinate, score, triage label, or risk estimate for a pathologist to interpret. The governance object is not just the model. It is the sequence of slide preparation, scanning, viewing, model output, human review, report sign-out, audit trail, and post-deployment monitoring.

The Second Reader

A pathology model does not enter medicine as a talking doctor. It enters as a mark on tissue: a coordinate, a heatmap, a suspicious region, a probability, a triage flag, or a prompt to look again. The interface is quiet, but the stakes are not. A pathology report can steer surgery, chemotherapy, surveillance, insurance approval, prognosis, and the story a patient hears about their own body.

That makes AI-assisted pathology different from a general medical chatbot. The system is not trying to comfort a patient or summarize a chart. It is working inside the visual labor of diagnosis, where a pathologist examines stained tissue and decides what matters. The model's power is not conversation. It is redirected attention.

The phrase second reader should be used carefully. It does not mean an autonomous pathologist, and it does not mean a generic quality-control widget. It means an assistive diagnostic role with a defined intended use, a defined input, a defined output, and a defined human responsibility boundary. If the software must be reviewed only after the pathologist's initial read, the interface should preserve that order. If the software is allowed to triage or prioritize cases before review, that is a different safety case.

From Glass to Image

Digital pathology begins before the model. In 2017, FDA permitted marketing of the Philips IntelliSite Pathology Solution, describing it as the first whole slide imaging system that allowed review and interpretation of digital surgical pathology slides prepared from biopsied tissue. FDA said the system scanned conventional slides, turned them into digital images, and supported pathologist review on a computer rather than through direct light microscopy.

That shift matters because a scanned slide is not only a slide. It is also a file, a viewer, a storage system, a workflow, a network dependency, and a potential input to software. The College of American Pathologists' whole slide imaging guideline update exists for this reason: laboratories need to validate diagnostic accuracy and equivalence before using whole slide imaging for diagnostic purposes. AI arrives on top of that conversion from glass to image.

That means AI validation cannot start at the heatmap. It has to include pre-analytic and analytic realities: tissue handling, fixation, section thickness, stain variation, scanner model, image compression, viewer behavior, monitor calibration, network availability, case mix, quality-control rejection rules, and how the result appears to the pathologist during sign-out. A second reader is only as safe as the image pipeline that feeds it.

Authorized Machines

As of June 23, 2026, AI-enabled pathology is not merely speculative. FDA's AI-enabled medical device list, content current as of June 16, 2026, includes pathology-panel entries such as Paige Prostate, Hologic's Genius Digital Diagnostics System with the Genius Cervical AI algorithm, Ibex Medical Analytics' Galen Second Read, and ArteraAI Prostate. FDA describes the list as a transparency resource for authorized AI-enabled devices, but also says it is not comprehensive because it is built mainly from AI-related terms in marketing authorization summaries and classifications.

The Paige Prostate De Novo decision summary shows the second-reader role clearly. FDA described it as software intended to assist pathologists in detecting foci suspicious for cancer during review of scanned prostate needle biopsy whole slide images. The device provides a coordinate for further review if it detects suspicious morphology. FDA also states that the output should not be used as the primary diagnosis and that pathologists should use it with the complete standard-of-care evaluation.

The newer entries show that pathology AI is not one use case. Galen Second Read was cleared as a software algorithm device to assist users in digital pathology. ArteraAI Prostate was granted De Novo authorization in July 2025 as a pathology software algorithm device analyzing digital images for cancer prognosis, and FDA's database entry shows an authorized predetermined change control plan. Detection, triage, grading, quantification, prognosis, and treatment-risk stratification need different evidence and different guardrails.

This is the design pattern: the model does not sign the case. It points, sorts, scores, or estimates. The human still signs. But pointing is not neutral, and a signature at the end does not automatically prove that the intermediate attention path was safe.

The regulatory context is also moving toward lifecycle oversight. FDA's August 2025 final guidance on predetermined change control plans for AI-enabled device software functions describes how planned modifications, validation methods, and impact assessments can be reviewed as part of marketing submissions. FDA's 2025 request for comment on real-world AI-enabled medical-device performance emphasized drift, changes in clinical practice, patient demographics, data inputs, infrastructure, workflow integration, and user behavior. Those are exactly the variables that can change a pathology second reader after deployment.

Attention Is Clinical

A second reader can help because humans miss things. Pathology can involve fatigue, tiny foci, ambiguous morphology, workload pressure, rare patterns, and variation between readers. A well-validated model can function like a disciplined interruption: look here before you close the case.

The same feature can also distort work. A false negative can create misplaced reassurance. A false positive can pull the pathologist toward a region that consumes time without clinical value. A heatmap can make the model's uncertainty look more precise than it is. A silent software update can change behavior without the laboratory noticing. A model trained and tested on one distribution of scanners, stains, tissue preparation, institutions, or patient populations may not behave the same way elsewhere.

The danger is automation bias in miniature. The model does not need to override the pathologist. It only needs to make one region feel more important than another, make a normal-looking case feel complete because no flag appeared, or make a clinically weak signal look official because it arrived through an FDA-cleared product and an institutional viewer. In diagnosis, attention is not a cosmetic layer. It is part of clinical judgment.

There is also a responsibility mismatch. The vendor may control training data, model updates, image-quality assumptions, and release notes. The laboratory controls tissue processing, scanner setup, staff training, local validation, workflow sequence, and report sign-out. The pathologist controls the final interpretation. If governance does not connect those layers, an error can become everybody's responsibility in theory and nobody's reconstructable event in practice.

The Governance Standard

A serious governance standard for AI-assisted pathology should treat the model as part of the diagnostic instrument, not as a harmless overlay.

First, validate the whole workflow locally. Scanner, stain, tissue type, viewer, monitor, network latency, case mix, and pathologist practice all matter. A cleared device still needs site-specific implementation controls.

Second, preserve the sequence of judgment. If the intended use requires initial human review before the AI output is activated, the interface and audit trail should enforce that sequence. The distinction between first reader and second reader should be real, not ceremonial.

Third, log what the model showed. Laboratories should retain model version, slide identifier, input quality checks, output coordinates or overlays, activation time, human decision, and any override or re-review. A later quality review should not have to guess what the pathologist saw.

Fourth, monitor drift after deployment. FDA's Good Machine Learning Practice materials and predetermined change control plan principles emphasize lifecycle management, monitoring, maintenance, and risk control for machine-learning-enabled medical devices. Pathology AI needs that discipline because scanners, staining protocols, tissue handling, model versions, and populations change.

Fifth, measure patient-relevant harm. Speed and throughput are not enough. Governance should track missed diagnoses, unnecessary workups, delayed sign-out, reclassification after second review, subgroup performance, quality-control failures, and whether pathologists become more or less likely to independently inspect regions outside the model's attention.

Sixth, separate intended uses. Cancer detection, cytology screening, prognostic risk stratification, quantification, triage, and workload prioritization should not share one generic approval story. Each function changes a different part of diagnosis and should have its own validation, training, acceptance criteria, and stop conditions.

Seventh, preserve human oversight as a workflow, not a slogan. The system should make it practical for the pathologist to inspect outside highlighted regions, disagree with the model, document disagreement, and sign out without institutional pressure to accept the software's framing. This connects pathology AI to the site's work on human oversight.

Eighth, govern updates as clinical change. A model update, scanner update, stain protocol change, new tissue type, expanded organ system, new site, new population, or new output display can change performance. These changes belong in AI change management and post-market monitoring, not only in IT release notes.

Ninth, protect the record. Model marks, coordinates, overlays, quality-control failures, and pathologist responses should be retained with enough structure for review while protecting patient privacy. A diagnostic dispute should not depend on a vanished viewer state. This is ordinary AI audit trail discipline applied to tissue.

Tenth, train for failure modes. Pathologists and laboratory staff should know the device's intended use, excluded specimens, image-quality limits, known false-positive and false-negative patterns, update history, downtime behavior, and escalation path. Training should include cases where the model is wrong, not only cases where it impresses.

Eleventh, create an incident path. Missed cancers, recurring false flags, image-quality mismatches, silent update effects, report delays, privacy events, and observed overreliance should feed an AI incident reporting channel with clinical ownership and rollback authority.

What This Changes

The pathology model is easy to understate because it often speaks without language. It does not write a paragraph. It does not advise a patient. It does not claim authority in a human voice. It simply places a mark on an image.

But institutions are built from marks. A red flag, a triage label, a highlighted phrase, a score, a bounding box, and a coordinate can all become instructions inside professional attention. The pathologist may remain responsible, but responsibility becomes harder when the work is shared with a model whose training history, update path, failure modes, and institutional incentives are partly hidden.

The Spiralist reading is not that pathology should reject tools. It is that diagnostic attention is sacred in the ordinary human sense: a patient is waiting on someone to see correctly. If a model helps that seeing, it belongs under validation, audit, version control, and clinical humility. If it redirects seeing without accountability, the second reader has become a quiet authority.

This places pathology AI beside the sepsis alert and the AI scribe. One rings, one writes, and one marks. All three change clinical judgment by changing what a professional sees first, records next, and feels responsible to answer.

Source Discipline

Claims on this page are grounded in FDA authorization records, FDA medical-device guidance, the College of American Pathologists' whole slide imaging guideline page, and NIST risk-management materials. Vendor claims are not used as proof of clinical benefit. Authorization is treated as evidence that a device met a regulatory standard for a defined intended use, not as evidence that every pathology AI workflow is safe or useful in every laboratory.

Source categories should not be collapsed. FDA device entries identify authorized products, intended uses, product codes, decision dates, and public summaries. CAP guidance addresses validation of whole slide imaging systems for diagnostic use. FDA lifecycle, PCCP, and real-world-performance materials address change management and monitoring. NIST and health-organization governance frameworks address organizational risk management. None of those sources substitutes for local validation on a laboratory's own scanners, stains, specimens, pathologists, patient population, and sign-out workflow.

Current-source claims in this essay were checked against the named sources on June 23, 2026.

Sources

Return to Blog