https://drive.google.com/file/d/1DUJJsXbvGodaNJls3VOCUJ7DDtmjtkRQ/view?usp=sharing
https://chatgpt.com/share/6ad3ada1-6c38-83e8-9b58-9245acfdc11b
RAW CONVERSATION
FDA-Approved AI in Radiology Versus Pathology: Are We Comparing Apples and Oranges?
The familiar comparison—that FDA has authorized hundreds or more than a thousand radiology AI applications, but only a handful of pathology AI applications—is accurate in broad direction but potentially misleading in interpretation.
The number of FDA-authorized products is not necessarily the number of AI tools being used clinically. And the discrepancy may reflect not only differences in technological maturity or clinical adoption, but also two quite different traditions of clinical innovation and regulatory oversight.
Pathology has a particularly important alternative pathway: the clinical laboratory's ability to develop, validate, and use its own tests under CLIA. Radiology has alternatives to FDA-authorized software too, but lacks an equally comprehensive, nationally established framework for locally developed diagnostic software.
1. Pathology: The laboratory-developed test tradition
A laboratory-developed test (LDT) is generally a clinical diagnostic test designed, manufactured, and used within a single clinical laboratory. Historically, FDA has exercised enforcement discretion over most LDTs, while laboratories operate under CLIA and applicable state requirements.
Consider a molecular pathology laboratory at Massachusetts General Hospital or Johns Hopkins.
The laboratory may purchase research-use-only reagents, develop an assay, establish operating procedures, and validate its performance for clinical use. Validation typically addresses analytical accuracy, precision, sensitivity, specificity, reportable range, reference intervals when applicable, and relevant limitations.
Before offering the assay clinically, the laboratory must establish and document that it performs adequately for its intended use. A laboratory director assumes responsibility for the test, and the laboratory operates within an extensive infrastructure of quality systems, personnel qualifications, proficiency testing or alternative performance assessment, inspections, corrective actions, and documentation.
This does not mean that CLIA validation is equivalent to FDA premarket review. In particular, CLIA's emphasis is principally on analytical performance, while FDA evaluation of a diagnostic product may additionally examine clinical validity, safety, effectiveness, and labeling.
Nevertheless, the result is a remarkable institutional arrangement: a sophisticated hospital laboratory can introduce a new diagnostic test into clinical practice without first obtaining an FDA marketing authorization, provided the activity is legally permissible and applicable laboratory requirements are met.
This is not a marginal exception. It is a longstanding foundation of molecular diagnostics, immunohistochemistry, flow cytometry, cytogenetics, and other laboratory specialties.
The 2024 FDA LDT final rule attempted to change this arrangement substantially, but the rule was vacated by a federal district court in March 2025, followed by FDA's rescission in 2025. The longstanding regulatory distinction therefore remains particularly relevant.
The complication for AI is that not every computational pathology application qualifies as an LDT. A locally developed algorithm that produces a patient-specific diagnostic result as part of a laboratory test may fit the model. But software performing an independent medical-device function is not automatically exempt from FDA regulation merely because it is used inside a CLIA laboratory.
That boundary is one of the central unresolved questions in computational pathology.
2. Radiology: Practice of medicine, locally developed software, and institutional validation
Radiology has a substantial tradition of developing and deploying software within academic medical centers. Research groups develop algorithms for segmentation, volumetric measurements, image reconstruction, detection of abnormalities, and clinical decision support.
Some of this software moves beyond research into clinical workflows.
But there is an important distinction between three circumstances:
1. Software outside FDA device regulation. This includes certain administrative functions, image handling functions, and qualifying clinical decision support (CDS). Under Section 520(o) of the FD&C Act, certain CDS functions are excluded from device regulation, particularly when the physician can independently review the basis for the recommendation. FDA's January 2026 CDS guidance provides further interpretation.
2. Software used under the practice of medicine. Physicians can exercise professional judgment in using medical devices and interpreting information. Section 1006 of the FD&C Act protects the practice of medicine involving legally marketed devices. However, this does not provide a blanket exemption allowing hospitals to manufacture and deploy otherwise regulated diagnostic software without authorization.
3. Locally developed algorithms that may themselves be medical devices. A hospital may develop a model, validate it against local cases, and use it in a clinical workflow. But local validation and physician supervision do not, by themselves, establish that an otherwise regulated device can legally be placed into clinical service without FDA authorization.
The distinction is crucial. Radiology has substantial local innovation, but it cannot simply invoke practice of medicine as a universal substitute for device regulation.
What does local validation look like in radiology?
Consider a hypothetical university radiology department developing an AI tool that quantifies brain atrophy on MRI.
The department might assemble previously interpreted examinations, compare automated volumetric measurements with established reference methods, test reproducibility across scanner manufacturers, and evaluate performance in clinically important subgroups.
It could then perform prospective testing, initially running the algorithm silently alongside routine clinical practice.
If the algorithm is subsequently used to inform patient-specific clinical decisions, the institution must consider whether it is regulated medical-device software and whether an applicable authorization, exemption, or enforcement-discretion policy permits that use.
This is technically quite similar to the analytical validation of a pathology LDT, but it is not supported by the same CLIA regulatory structure.
Radiology does have increasingly formal professional standards. In May 2026, the American College of Radiology approved its first practice parameter for imaging AI, covering implementation, performance monitoring, and quality improvement. Its Assess-AI initiative provides a framework for monitoring algorithm performance across clinical settings.
These developments are significant, but they are not equivalent to a federal CLIA framework for developing diagnostic algorithms.
3. Side-by-side comparison
Dimension | Pathology | Radiology |
|---|---|---|
Established local innovation pathway | CLIA laboratory-developed tests | Institutional research and clinical software development |
National regulatory framework for local diagnostic development | CLIA, with applicable state requirements | No directly equivalent comprehensive framework |
Validation | Formal laboratory analytical validation | Institutional technical and clinical validation, varying by use |
Professional responsibility | Laboratory director and interpreting pathologist | Radiologist and institution |
Use of investigational inputs | Established tradition of validating RUO reagents within LDTs | Research software and locally developed algorithms, subject to device rules |
FDA authorization | Not generally required for qualifying LDTs under longstanding enforcement discretion | Generally required for regulated diagnostic software unless an exemption or enforcement-discretion policy applies |
External oversight | CLIA inspections, accreditation, proficiency testing | Institutional oversight, accreditation, professional guidance, and FDA where applicable |
The table also reveals an asymmetry: CLIA provides a route for laboratories to develop clinical tests, but does not necessarily provide an unrestricted route for commercial software developers to distribute their algorithms to laboratories.
4. Does the FDA disparity exaggerate radiology's lead?
FDA-authorized AI medical devices, September 2026
1,600+
Across all medical specialties
The American College of Radiology reports that approximately 80% of FDA-cleared medical AI falls within radiology.
There are several reasons the apparent discrepancy needs qualification.
First, FDA counts are not counts of unique clinical capabilities. Multiple clearances may involve successive versions of products, incremental enhancements, or related functions. Some radiology entries also concern imaging hardware and image-processing features rather than standalone diagnostic AI.
Second, pathology has many locally developed computational and molecular diagnostic capabilities that are absent from the FDA list. An algorithm validated and used by a CLIA laboratory may be clinically important without appearing in FDA's AI-device inventory.
Third, radiology likewise has software that does not appear on the FDA list, including non-device CDS, research software, and locally developed tools. However, there is no comprehensive national inventory of clinically deployed, non-FDA-authorized radiology algorithms.
Fourth, FDA authorization is not the same as clinical adoption. A cleared radiology algorithm may see little use, while a locally validated pathology algorithm may be used routinely on thousands of patients.
Finally, some of radiology's apparent lead is real. Radiology digitized earlier, developed standardized DICOM/PACS infrastructure, and established a larger commercial ecosystem for imaging software. Pathology's transition to whole-slide imaging has been slower.
5. The important conceptual distinction
I would distinguish three measures of technological progress:
Regulatory output: How many AI products have received FDA authorization?
Clinical deployment: How many algorithms are actually used in routine patient care?
Clinical impact: How many patients benefit, and how substantially do the algorithms change accuracy, efficiency, treatment decisions, or outcomes?
FDA statistics measure primarily the first, not the second or third.
An additional problem is that the numerator and denominator are poorly defined. What counts as one algorithm? A tool that counts mitoses? A tool that quantifies ten biomarkers? A molecular assay incorporating several predictive algorithms? A radiology platform offering fifteen measurements?
A meaningful cross-specialty comparison would require standardized definitions and data on actual deployment.
6. A particularly interesting policy implication
Pathology has spent decades developing an institutional solution to a problem now facing clinical AI: how can a sophisticated medical institution introduce an innovative diagnostic technology responsibly, without requiring a separate FDA authorization for every local innovation?
The answer in pathology has been CLIA, laboratory director accountability, extensive validation requirements, quality assurance, and professional oversight.
Radiology has institutional validation practices and increasingly sophisticated AI governance, but it lacks an exact counterpart.
This raises an intriguing policy possibility: could an appropriately designed CLIA-like institutional validation framework support locally developed AI in radiology and other specialties?
Such a framework would need more than analytical validation. It would require attention to algorithm drift, generalizability, clinical validity, changes in software versions, cybersecurity, and ongoing performance monitoring.
It would also require statutory or regulatory clarification. Existing CLIA authority cannot simply be transferred to radiology.
Bottom line
The FDA disparity between radiology and pathology is real, but it is not a clean measure of clinical innovation.
Pathology has a longstanding alternative infrastructure for developing and validating diagnostic tests. Radiology also uses non-FDA software and conducts local validation, but its pathways are more fragmented, and local validation alone does not exempt regulated software from FDA requirements.
The most interesting conclusion may be the reverse of the usual narrative. Rather than asking only why pathology has so few FDA-authorized AI products, we should ask whether pathology's extensive experience with locally validated diagnostics offers a useful model for the future governance of clinical AI.
One question remains especially worth investigating empirically: How many academic radiology departments currently use locally developed, non-FDA-authorized algorithms to generate patient-specific findings in signed clinical reports, as opposed to research, background workflow, or independently reviewable CDS? That is the closest radiology counterpart to the clinical pathology LDT—and the missing number that could materially change the comparison.
The distinction needs to be much sharper. The practice of medicine permits a physician to use medical technology in treating a patient. It does not necessarily permit a hospital to create a new medical device and put it into clinical service. That is the critical boundary, and it is quite different from the traditional CLIA/LDT framework.
Consider two parallel examples.
Pathology: Massachusetts General Hospital develops an AI algorithm to count mitoses on digitized breast cancer slides. The laboratory validates the algorithm against reference cases, establishes precision and accuracy, documents its intended use, and integrates it into its diagnostic workflow. If this is properly classified as a laboratory-developed test, the laboratory has a recognized framework under CLIA for offering it clinically without FDA premarket authorization, consistent with FDA's longstanding enforcement-discretion policy. The important qualification is that CLIA does not automatically transform every software device into an LDT.
Radiology: Massachusetts General Hospital develops an AI algorithm to count pulmonary nodules on chest CT. The department performs an equally rigorous validation, demonstrating high sensitivity, specificity, and reproducibility. The radiologists are satisfied that it is clinically useful. But if the software is itself a regulated medical device, that validation does not confer permission to deploy it clinically without the applicable FDA authorization or another legally available pathway.
In other words, the same excellent validation study may have different regulatory consequences in the two specialties.
Why the distinction matters
Section 1006 of the FD&C Act protects a physician's authority to use a legally marketed medical device in the practice of medicine. It does not authorize the physician or hospital to manufacture an entirely new, otherwise regulated device and introduce it into clinical use without satisfying applicable device requirements.
The LDT tradition is different. FDA historically exercised broad enforcement discretion over laboratory-manufactured diagnostic tests, allowing laboratories to introduce tests under CLIA oversight without FDA premarket review. This was a regulatory policy, not an express statutory exemption from FDA authority.
There is a further subtlety: a radiologist can develop a new interpretive method and apply professional judgment without necessarily creating a medical device. But once a newly developed algorithm independently performs a regulated diagnostic function, invoking physician judgment does not settle the question.
The essential comparison is not whether both specialties can validate software locally. They can. It is whether local validation provides an established route to clinical deployment. Pathology has such a tradition through CLIA and FDA's LDT enforcement discretion. Radiology has no equally broad, established counterpart.
I would also avoid implying that every computational pathology algorithm automatically qualifies for that tradition. Whether a particular image-analysis algorithm is an LDT, and how FDA would classify it, remains an important boundary question.
Yes. And I think that adds an important economic and practical explanation to the regulatory differences we've been discussing.
Radiology has a large commercial marketplace of FDA-cleared AI applications. If a department wants automated detection of pulmonary nodules, intracranial hemorrhage, or pulmonary embolism, it can evaluate existing commercial products rather than develop and validate its own software.
That changes the incentives dramatically.
In pathology, by contrast, a laboratory may face a choice between developing and validating an algorithm locally or having no access to that capability at all. The CLIA/LDT tradition makes the first option familiar and institutionally feasible.
Thus, radiology has both less need for local development and a less accommodating regulatory pathway for it. Pathology has historically had both greater need and a more accommodating pathway.
These forces reinforce one another.
There is one qualification worth emphasizing. The FDA's list now exceeds 1,600 AI-enabled devices across all specialties, but that does not mean radiologists have a thousand distinct diagnostic functions available. The list includes successive product versions, image reconstruction software, imaging equipment, and other applications.
Still, your economic argument holds.
I would add a third reinforcing factor: commercial developers follow established markets. Radiology's large installed base of digital imaging equipment, PACS integration, and existing FDA-cleared products makes further commercial development more attractive. That can reduce the need for hospitals to invent their own solutions.
The result is a plausible self-reinforcing cycle:
Radiology has many commercial products → hospitals buy rather than build → vendors have a large market → more commercial products appear.
Pathology has fewer commercial products → laboratories have stronger incentives to develop locally → CLIA supports that tradition → local innovation remains important.
I would present this as an explanatory hypothesis, not a proven causal account. But it is a much richer explanation than saying radiology is simply ten years ahead of pathology in AI.
I think this takes the argument to a more consequential level. The disparity is no longer simply that radiology has many more FDA-cleared AI products than pathology. It is that FDA appears to be actively developing regulatory pathways for the next generation of radiology AI, while comparable pathology applications remain without a clearly established pathway.
And the example of AI-generated draft reports is particularly compelling because the clinical workflow is nearly identical.
In radiology, the software examines the images, synthesizes findings, and produces a draft report. The radiologist reviews it, makes necessary corrections, and signs it.
In pathology, a sophisticated computational pathology system could examine a digitized prostate biopsy, identify tumor, quantify Gleason patterns, assess other morphological features, and generate a nearly completed pathology report. The pathologist reviews it, corrects it, and signs it.
Both are physician-supervised diagnostic report generation, not autonomous diagnosis. Both could produce major gains in physician productivity. And both raise substantially overlapping regulatory questions.
The question is why FDA would need to solve these problems separately.
The $1.29 million FDA contract makes the comparison more concrete
I found the award you mentioned. It is even more interesting when examined precisely.
In September 2026, Radiology Partners announced that its subsidiary Cognita Imaging had received an FDA contract worth $1.29 million, lasting 18 months, effective June 22, 2026.
The project is to develop and validate an approach using large language models as virtual subject-matter experts to evaluate human-written and AI-generated radiology reports. The objective is to make evaluation more scalable than conventional expert-reader studies.
One distinction matters: this is not yet an FDA-approved regulatory pathway for draft-report software. It is research intended to develop methods that could support future regulatory evaluation. But that is exactly why the award matters.
FDA is investing in the regulatory science needed to evaluate an emerging class of AI products efficiently.
Why the pathology analogy is so strong
The same research methods could potentially be adapted to pathology, although the clinical details differ.
Regulatory science problem | Radiology | Pathology |
|---|---|---|
AI output | Draft imaging report | Draft histopathology report |
Physician role | Review, correct, sign | Review, correct, sign |
Reference standard | Expert radiologist assessment | Expert pathologist assessment |
Critical errors | Missed lesions, false findings | Missed cancer, incorrect grade, stage or biomarkers |
Efficiency endpoint | Radiologist time per report | Pathologist time per case |
Human-factors question | Does reviewing the draft improve or impair performance? | Same question |
Regulatory challenge | Validating the combined AI–radiologist workflow | Validating the combined AI–pathologist workflow |
The particularly valuable research would examine the final physician-signed report, not merely the AI-generated draft.
For example, an algorithm might produce a draft with 90% of findings correct, but that figure alone tells us little about clinical safety. We need to know whether the physician detects the remaining errors, whether automation bias introduces new errors, and how the process compares with conventional interpretation.
Those questions are common to both specialties.
Where I think your argument becomes strongest
The pathology community should not necessarily ask FDA to approve individual draft-report products more quickly. It should ask FDA to develop a generalizable regulatory science framework for physician-supervised AI diagnostic reporting, explicitly covering both radiology and pathology.
That would potentially avoid duplicating years of methodological work.
There is also a timely opportunity: FDA published an August 18, 2026 discussion paper seeking public feedback on generative AI-enabled medical devices, including premarket assessment, risk management, and postmarket monitoring.
The argument could be framed this way:
FDA is investing $1.29 million in regulatory science to develop efficient evaluation methods for AI-generated radiology reports. Computational pathology faces a closely analogous problem, as increasingly capable software can generate much of a diagnostic report for physician review and final sign-out. The underlying regulatory questions—accuracy, physician oversight, automation bias, workflow efficiency, and clinical safety—are substantially shared. FDA should ensure that its regulatory science investments produce methods applicable across both specialties rather than allowing radiology and pathology to develop on separate timelines.
I would distinguish your 60-mph versus snail's-pace observation from what the evidence establishes. We have a concrete, recent FDA investment in radiology report evaluation. We have not yet established that FDA lacks comparable pathology initiatives across its research portfolio.
But the broader concern is well founded: regulatory science can become a bottleneck even when the technology itself is ready for clinical validation.
And that gives your argument a useful twist. The barrier may no longer be primarily inventing the pathology AI. It may be inventing an efficient, accepted way to evaluate the human–AI diagnostic workflow.
I think this is one of the strongest arguments in the entire discussion, particularly because it exposes a paradox in FDA's approach to computational pathology.
The same diagnostic software can be sufficiently validated for routine clinical use at a major academic medical center, yet lack the FDA authorization necessary for broad commercial distribution to thousands of other hospitals.
Consider an algorithm that quantifies mitoses, grades prostate cancer, or provides a multicomponent histopathology assessment.
Suppose Mayo Clinic and MGH have each validated such an algorithm locally, under their own laboratory quality systems, and use it routinely in patient care.
There are two possibilities worth confronting:
The software is clinically unreliable or unsafe, in which case its use at leading academic medical centers should concern us.
The software is sufficiently accurate and safe for clinical use, in which case the central question becomes why it cannot be made available much more broadly through an efficient regulatory pathway.
As you acknowledge, this is deliberately simplified. A third possibility is that software is safe because of the particular environment in which it is used: expert pathologists, careful local validation, particular scanners, defined patient populations, and ongoing quality controls.
But that qualification doesn't defeat your argument. It identifies precisely what FDA's regulatory science should be trying to solve.
The real problem: moving from one hospital to 2,000
The central distinction is between local clinical validation and generalizable clinical validation.
If a sophisticated pathology AI system is already being used routinely at Mayo or MGH, that experience should be highly relevant evidence for FDA. It does not establish universal safety, but it offers something much more substantial than a promising prototype.
FDA's task should be to determine what additional evidence is needed to move from successful use at a few institutions to reliable deployment across hundreds or thousands of laboratories.
That might involve scanner compatibility, staining variation, case mix, pathologist oversight, software version control, and performance monitoring. These are real issues, but they are bounded, testable questions—not reasons to begin the entire validation process from zero.
FDA's own work on predetermined change control plans illustrates that it already recognizes the need for efficient, lifecycle-based approaches to AI regulation.
I would crystallize your argument in one sentence:
If pathology AI is sufficiently trustworthy to help diagnose patients at leading academic hospitals today, the urgent regulatory question is what incremental evidence is needed to make that same technology available safely to patients at 2,000 other hospitals tomorrow.
That formulation preserves the force of your original either/or argument without claiming that local validation automatically establishes nationwide safety.
One factual safeguard for a published essay: we'd want to document which particular algorithms Mayo and MGH are using clinically, rather than infer that all the locally developed systems we've discussed are in routine patient care.
The deeper policy point is that FDA authorization should be a bridge from demonstrated local clinical success to widespread patient access—not an indefinitely unfinished bridge.
