[Chat GPT 5.5 Guest Author]
##
Digital Pathology AI Moves from “Cool Demo” to Clinical Infrastructure:
The New DPA Recommendation Statement
A new recommendation statement from the Digital Pathology Association arrives at exactly the right time.
For several years, computational pathology has had the flavor of a rapidly advancing frontier: striking images, impressive AI heat maps, strong performance statistics, and a growing list of FDA-authorized tools. But the practical question has always been harder: how does computational pathology actually become part of clinical laboratory medicine?
That question is no longer theoretical.
Digital pathology adoption is expanding. AI tools for H&E slides, biomarker scoring, risk stratification, quality control, triage, and workflow support are moving from research into clinical deployment. The AMA CPT process appears, at least for H&E-based computational pathology, to be channeling many of these services into Category III codes. FDA authorizations may begin to accelerate. Payers will increasingly be asked to distinguish between “AI that works in a paper” and “AI that deserves clinical adoption and reimbursement.”
Into this setting comes a timely and unusually comprehensive review and recommendation statement from Silberman and colleagues, published in AI in Precision Oncology under the auspices of the Digital Pathology Association.
Full reference:
Silberman N, Parwani A, McClintock DS, Lujan G, Gondim DD, Garcia C, Hanna MG, Pantanowitz L, Lennerz JK, Showalter T, Gerrard P. Recommendation Statement for the Validation, Implementation, and Clinical Application of Artificial Intelligence Within a Clinical Laboratory from the Digital Pathology Association. AI in Precision Oncology. 2026;00(00). doi:10.1177/2993091X261455975.
Full link: https://journals.sagepub.com/doi/10.1177/2993091X261455975
Overview at Linked In by Lennerz: here.
Not Just “Does the AI Work?”
A central virtue of the Silberman paper is that it does not treat computational pathology as a single problem. Instead, it separates several layers that are often blurred together.
First, there is the digital pathology infrastructure: slide scanners, image viewers, displays, laboratory information system interfaces, image management systems, and workflow integration. Second, there is the AI algorithm itself. Third, there is the clinical setting in which the AI is used. Fourth, there is the question of whether the AI output has clinical utility.
That distinction matters. A lab might have a validated scanner and viewer for primary diagnosis, but that does not automatically validate a new AI model. Conversely, an AI model might perform well on a certain scanner or staining protocol, but that does not automatically mean it can be moved to another scanner, another lab, or another patient population.
The paper repeatedly emphasizes this point: digital pathology infrastructure and AI algorithms require distinct validation. That sounds obvious once stated, but it is a critical operational point for laboratories, developers, regulators, and payers.
Pathologist Oversight Is Not a Footnote
The article is also clear that clinical AI in pathology remains under qualified pathologist oversight. This is not merely a professional turf statement. It reflects how pathology services are delivered, validated, signed out, and defended clinically.
Most computational pathology tools today are not replacing pathologists. They are augmenting them. They may find tumor regions, quantify biomarkers, identify mitoses, triage cases, improve consistency, or help extract prognostic information from morphology. But the clinical result still has to live inside a diagnostic workflow.
This has implications for coding and payment. Some AI services may be best understood as part of the physician pathology service: a tool used by the pathologist in reaching or supporting a diagnosis. Other AI services, especially if they produce a stand-alone result without physician interpretive work, may look more like clinical diagnostic laboratory tests. Silberman et al. directly recognize that AI applications in digital pathology may fall into either category depending on use case.
This is exactly where the AMA coding process, CLIA, FDA, and Medicare policy are likely to collide.
Analytical Validity, Clinical Validity, Clinical Utility
The article usefully frames evidence around three concepts familiar from laboratory medicine but sometimes underdeveloped in AI discussions.
Analytical validation asks whether the system reliably and accurately measures what it claims to measure under defined conditions. For computational pathology, this includes such issues as scanner dependence, staining variation, tissue adequacy, artifacts, batch effects, repeatability, reproducibility, and quality control.
Clinical validation asks whether the output is clinically meaningful in the intended population and setting. For example, does a score predict recurrence, response, diagnosis, or another clinically relevant state with adequate performance?
Clinical utility asks whether using the test improves patient outcomes, either directly or through a convincing chain of evidence.
That third step is often the hard one. A model can be analytically impressive and clinically associated with outcome, yet still fail to change care in a meaningful way. Conversely, a tool that simply improves turnaround time, consistency, access to expertise, or case triage may have real utility even if it does not look like a traditional “new biomarker.”
This is one of the paper’s most important contributions. It broadens the conversation beyond AUCs and validation cohorts and forces the field to address how AI changes care.
Three Types of Clinical Use
The DPA statement identifies three broad types of appropriate clinical use for AI in digital pathology.
First, AI may replace an existing invasive, expensive, slow, or otherwise burdensome test when adequately validated. This is the most straightforward payer narrative: the AI does the job of something already recognized as clinically useful, but faster, cheaper, or more scalably.
Second, AI may augment pathologist assessment of verifiable tasks. Examples include quantifying features, improving consistency in scoring, detecting rare events, or reducing variability. These applications may be less glamorous but could be among the earliest to diffuse broadly.
Third, AI may predict outcomes from features that pathologists cannot readily assess by eye. This is the most ambitious category. It includes morphology-based prognostic or predictive tools that infer risk, treatment response, or other clinically meaningful outcomes from H&E or other images.
For Medicare and other payers, these three categories are likely to behave very differently. Replacing an existing test may be easier to explain economically. Augmenting pathologist assessment may be easier to defend clinically but harder to price. Predicting outcomes from invisible morphology may be scientifically exciting but may require the most careful evidence development.
Why Category III Coding Matters
The Category III direction emerging from the AMA process is important because it signals that the field is still in an evidence-collection phase. Category III codes do not mean a service is unimportant. Quite the opposite: they often mark services that are new, rapidly evolving, and significant enough to track.
For computational pathology, Category III coding may create a paradox. These services may be technologically sophisticated, capital-intensive, and potentially valuable, but the Category III pathway does not itself guarantee national pricing, coverage, or favorable payment. In practice, the evidence package may become as important as the code.
That is why the Silberman paper matters. It provides a language for developers and laboratories to explain not merely what their AI does, but how it was validated, where it is generalizable, what role the pathologist plays, what limitations apply, and how clinical utility should be established.
In a world of Category III codes, the evidence framework may become the product.
The FDA Piece
The article also discusses the regulatory boundary around software as a medical device, including AI tools that diagnose, treat, or inform clinical decisions. Some AI pathology tools clearly enter FDA territory, such as cancer detection, automated grading, or companion diagnostic selection. Other software functions, such as storage, transfer, viewing, or image management without interpretation, may not be medical devices.
The paper also reflects the post-2025 legal environment around laboratory-developed tests, noting the Texas federal court decision vacating FDA’s attempt to regulate LDTs as medical devices. The key distinction is between the device and the professional laboratory service using the device. This issue will remain fluid, but it is directly relevant to pathology AI.
Computational pathology may therefore develop through multiple channels: FDA-authorized software, CLIA laboratory services, pathologist-supervised workflows, and payer-defined evidence requirements. That is not unusual in diagnostics, but it does mean the field will not have one single regulatory or reimbursement pathway.
Access to Expertise: The Quietly Powerful Argument
One of the most interesting parts of the article is its attention to access. Digital pathology is not only about algorithmic performance. It also enables remote review, subspecialty consultation, worklist triage, and potentially better access for patients served by rural hospitals, Critical Access Hospitals, Rural Health Clinics, and Federally Qualified Health Centers.
This is a strong policy argument. Pathology expertise is unevenly distributed. Cancer diagnosis is increasingly complex. Precision oncology depends on correct tissue diagnosis, biomarker interpretation, and integration of morphology with molecular data. Digital pathology can move expertise to the slide rather than moving the patient, the specimen, or the specialist.
AI may strengthen this model by triaging urgent cases, flagging suspicious slides, helping detect cancer, or supporting standardized scoring. In that sense, computational pathology is not simply an “add-on” to pathology. It may become part of the national infrastructure for distributing diagnostic expertise.
Multimedia Pathology Reports?
A smaller but fascinating discussion in the paper concerns patient and clinician access to pathology images. Radiology has long allowed clinicians to show patients their own images. Pathology, by contrast, has remained largely text-based, partly because glass slides are fragile, specialized, and hard to view outside the lab.
Digital pathology changes that. Images can accompany reports. Clinicians can show patients the tissue basis for diagnosis. Pathology may become more visible, more understandable, and more directly connected to patient decision-making.
This may sound secondary compared with AI scoring or prognostic modeling, but it could have cultural importance. Pathology has often been central to diagnosis but invisible to patients. Digital pathology may make the pathologist’s work more concrete and accessible.
Sidebar:Ten Non-Obvious Takeaways from the DPA Statement
Scanner validation and AI validation are not the same thing. A validated digital pathology platform does not automatically validate every AI model that runs on it.
Interscanner generalizability is now a first-class issue. The paper treats scanner-to-scanner concordance as a serious validation problem, not a minor technical nuisance.
Reliability matters separately from accuracy. An algorithm that is “right on average” may still be clinically problematic if it is unstable under real-world staining, scanning, tissue, or batch conditions.
Minimum tissue requirements and artifact limitations should be explicit. AI outputs should not be treated as magic numbers generated from any image under any condition.
The paper gives payers a framework, not just laboratories. Clinical utility, cost-effectiveness, generalizability, and quality alignment are presented as payer-relevant concepts.
AI can be useful even when it does not discover a new diagnosis. Triage, workflow prioritization, consistency, and access to expertise may be clinically meaningful uses.
The article recognizes that AI pathology services may fall on either side of the physician-service/lab-test divide. That distinction will matter for CPT, Medicare payment, and coverage policy.
Patient-facing pathology images are treated as part of the future. Digital pathology may eventually make pathology reports more visual, understandable, and clinically engaging.
The strongest AI use cases may have very different reimbursement stories. Replacing an existing expensive test, augmenting pathologist scoring, and predicting outcomes from invisible morphology are not the same payment problem.
The paper quietly shifts the field from “AI accuracy” to “AI governance.” The key questions become validation, limitations, oversight, monitoring, utility, and reproducibility.
Bottom Line
The Silberman/DPA recommendation statement is important because it treats computational pathology as clinical laboratory medicine, not as a software demonstration.
That framing is exactly what the field needs now. As CPT Category III codes appear, FDA authorizations accumulate, and laboratories begin deploying H&E-based and other AI tools, the central questions will be practical: What was validated? On which scanners? In which populations? Under whose oversight? With what limitations? For what clinical purpose? With what evidence of utility?
Computational pathology is likely to become an important layer on top of digital pathology. But it will not be adopted simply because it is accurate in a publication. It will be adopted when laboratories, pathologists, regulators, payers, and clinicians can trust that it is validated, reliable, clinically meaningful, and useful in real-world care.
That is the contribution of the Silberman paper. It gives the field a map for moving from enthusiasm to infrastructure.
###
###
Silberman et al explicitly discuss (pdf p. 10) the AMA's classification of AI medical software, aka AMA CPT "Appendix S," recently under extensive revision. Discuss Silberman vs AMA CPT as a second sidebar.
###
###
Sidebar 2:
Where Silberman and AMA Appendix S Overlap —
and Where They May Pull Apart
Silberman et al. and the AMA’s May 2026 Option D revision to Appendix S are not directly linked documents. They have different purposes.
- Silberman is a Digital Pathology Association recommendation statement about validation, implementation, clinical utility, and laboratory deployment of AI in digital pathology.
- Appendix S is a CPT taxonomy for describing AI services in code descriptors.
Still, the two documents are converging on the same fast-moving frontier: how medicine should name, validate, supervise, and eventually pay for clinical AI.
There is real overlap.
- Both documents reject the idea that “AI” by itself is a meaningful descriptor.
- Both ask what the software actually does.
- Both distinguish between AI that merely supports a clinician and AI that produces a clinically meaningful score, classification, risk estimate, diagnosis, or management recommendation.
- Both recognize that validity and clinical meaningfulness matter. Both are trying to move the field from marketing language to operational language.
But the center of gravity is different.
Appendix S is fundamentally a coding taxonomy. It asks whether the software output is assistive, augmentative, or autonomous. Assistive software provides clinically relevant information without deriving a score, index, classification, interpretation, or conclusion. Augmentative software derives a clinically meaningful parameter, such as a risk score, index, classification, or prognostic metric. Autonomous software goes further: it derives parameters and independently generates clinically meaningful interpretations, diagnoses, or management recommendations.
That taxonomy is useful, but it has a strong radiology flavor. Much of the conceptual machinery seems designed around imaging services in which software analyzes an image, generates an output, and that output either supports, modifies, or substitutes for some portion of physician interpretation. The examples in the Appendix S table — algorithmic ECG assessment, coronary FFR from imaging, retinal imaging — are not pathology examples. They are closer to the familiar world of radiology, cardiology, and ophthalmology AI.
Digital-computational pathology is similar, but not identical.
In pathology, the “image” is not simply an acquired image. It is the endpoint of a laboratory process. Tissue fixation, embedding, cutting, staining, coverslipping, scanning, image management, and algorithmic analysis are all part of the pipeline. A whole-slide image is not just a picture; it is a laboratory specimen transformed into a computational object. Silberman therefore emphasizes issues that Appendix S can easily understate: separate validation of scanning infrastructure and AI algorithms, scanner-to-scanner generalizability, staining variation, batch effects, minimum tissue requirements, artifacts, and quality control failures.
That is one potential conflict. Appendix S classifies the software output. Silberman asks whether the entire digital pathology system is valid in the laboratory setting where it is used. In pathology, the output category is only one piece of the problem.
A second difference is clinical utility. Appendix S says augmentative and autonomous outputs should be clinically meaningful and contribute to patient management. Silberman goes further into the laboratory-medicine tradition: analytical validation, clinical validation, and clinical utility are distinct concepts. A computational pathology score may be technically reproducible and clinically associated with outcome, yet still raise a separate question: does using it improve care, directly or through a convincing chain of evidence?
A third difference concerns the role of the physician. Appendix S classifies AI partly by whether physician or QHP interpretation is required. Silberman is more emphatic that clinical AI in pathology should remain under qualified pathologist oversight, even when the software provides information that may not require traditional pathologist interpretation. This reflects the reality that pathology is not simply image reading. It is laboratory medicine, specimen adequacy, diagnosis, integration of clinical facts, recognition of unexpected disease, and responsibility for the final report.
A fourth difference is that Appendix S’s autonomous category may be awkward in surgical pathology. In many radiology-style examples, one can imagine an AI system making a finding and triggering an action. In anatomic pathology, the pathologist’s report often remains the clinical instrument. Even a powerful H&E algorithm that predicts recurrence risk, response to therapy, tumor subtype, or biomarker status may still be embedded in a pathologist-signed laboratory report. Calling such a system “autonomous” because it generates a clinically meaningful conclusion may obscure the laboratory and professional structure in which the result is delivered.
A fifth difference is reimbursement. Appendix S is built for CPT descriptors, and it notes that physician work related to augmentative software may often be captured by existing codes. That may be sensible in many settings, but it creates an unresolved problem for computational pathology. If an H&E-based AI tool requires costly slide digitization, image storage, software licensing, quality systems, validation studies, monitoring, and data infrastructure, but the resulting output is treated as merely a data element inside an existing physician service, the coding taxonomy may fail to recognize the new laboratory resource costs.
This is especially important if H&E-based computational pathology is largely routed through Category III codes. Category III coding may track innovation but does not itself solve coverage, pricing, or valuation. For digital pathology AI, the cost structure may be closer to a laboratory platform than to a small interpretive aid.
The most useful reconciliation may be this: Appendix S can describe the software output, while Silberman can describe the clinical-laboratory evidentiary framework around that output.
Thus, a computational pathology service might be “augmentative” under Appendix S because it produces a prognostic risk score. But Silberman would still ask: Was the scanner validated? Was the algorithm validated on that scanner? Is the result robust to staining and tissue artifacts? What are the minimum tissue requirements? Was the score clinically validated in the intended-use population? Does it change patient management? Is there evidence of clinical utility? Who oversees the result? How is performance monitored over time?
For digital pathology, these are not secondary questions. They are the substance of the service.
The deeper issue is that Appendix S is a taxonomy of AI outputs, while computational pathology is a laboratory system.
That distinction may become increasingly important as FDA authorizations, Category III codes, and payer decisions accumulate. If CPT language remains too closely modeled on radiology-style image interpretation, it may miss the orthogonal problems that make computational pathology distinctive: specimen processing, scanner dependence, laboratory validation, image infrastructure, pathologist oversight, and clinical utility as understood in diagnostics.
In that sense, Silberman et al. do not contradict Appendix S. They complete it — or at least they show what Appendix S leaves unfinished for pathology.
