I've argued that the billion-dollar app in digital pathology will be an FDA-cleared app for draft report preparation. (The pathologist then supervises the slide results and confirms/edits the report). (My blog & PDF white paper.)
- Critical news 1: FDA has already starting approving some specific apps for draft reports in radiology.
- Critical news 2: FDA gave a $1M contract research award for how it should assess and approve report-generating apps in radiology. Worth tracking because: pathology won't be too far behind.
Report below by Chat GPT 6.
###
FDA Funds Research on AI Generated Radiology Reports and Pathology Should Watch
I’ve argued that the billion-dollar app in digital pathology will be an FDA-cleared application for draft report preparation. The software analyzes the slides and prepares a draft; the pathologist reviews the findings, confirms or edits the report, and signs it out. That is the thesis of my September blog and accompanying white paper. [1]
Radiology is already moving in this direction, including FDA clearance of specific applications with automated reporting capabilities. Now comes another development worth tracking: a $1.29 million FDA research contract to study how AI-generated radiology reports should be evaluated. My thesis is that radiology offers a reliable preview of changes that will reach pathology after a lag. This award addresses a problem both specialties will need to solve.
What FDA is funding
AuntMinnie reported the award on September 16, 2026, followed by Radiology Business on September 17. The recipient is Cognita Imaging, a subsidiary of Mosaic Clinical Technologies, which is itself owned by Radiology Partners.
Although the Radiology Business headline calls it a grant, the company announcement describes an 18-month research contract, effective June 22, 2026. [2–4]
The project is titled “Virtual Subject Matter Expert Agents for Radiology Report Evaluation.” Akshay Chaudhari, a Cognita co-founder and Stanford associate professor, is the principal investigator; Cognita CEO and co-founder Louis Blankemeier is co-investigator. The contract was awarded through FDA’s Broad Agency Announcement program for regulatory-science research. FDA describes that program as a way to bring outside expertise and infrastructure to problems in product evaluation and postmarket surveillance. [4,5]
The distinction matters: this award supports research on evaluation methods. It does not clear a Cognita reporting product, establish a new regulatory pathway, or announce FDA’s final standards for generative reporting.
- Would or could a separate grant and report on validation methods for pathology auto-reports (draft reports) ignite a tsunami of investment and progress in digital pathology?
- In radiology, see May et al. 2026 here and Beger here.
A jury of language models
Cognita’s proposed method uses several large language models to assess reports, comparing their judgments instead of relying on a single AI evaluator. After developing and validating this “Multi LLMs-as-a-jury” framework, the researchers plan to apply it to approximately one million patient examinations across different patient populations, clinical settings, equipment, and diseases. Radiologists will adjudicate clinically important disagreements. [2–4]
A particularly useful feature is that disagreement will not automatically mean the generated report is wrong. The problem could lie in that report, the AI evaluator, or the original radiologist’s report. The project will also reconstruct smaller validation cohorts from the large dataset to examine what smaller studies miss. Deliverables to FDA include software code, practical guidance, discrepancy studies, and analyses of the findings and limitations. [4]
This is an important change in the evaluation problem. A report can contain multiple findings, measurements, interpretations, and recommendations. Fluent prose can conceal a consequential omission or contradiction. The research therefore asks whether automated evaluation can extend expert review to a much larger and more varied case population. Agreement among several models, by itself, cannot establish clinical truth; shared errors remain possible. The quality of the evaluator must be tested along with the quality of the reporting system.
The broader research behind the award
The announcement points to earlier work on GREEN, an open-source radiology-report evaluator. Its researchers designed it to identify and explain clinically significant errors, addressing the limitations of text-similarity measures that may reward wording without adequately measuring factual correctness. The FDA contract builds on this direction by studying multiple evaluators and large-scale clinical variation. [4,6]
For pathology, the corresponding questions are easy to imagine. Did a draft assign the finding to the correct prostate core? Did it omit a small malignant focus, misstate grade, or confuse tumor extent across specimens? These are questions about diagnostic content and its faithful assembly into a report. A future evaluation framework will need to test those relationships.
 |
| click to enlarge; see my sept 1 blog on the left, the pathologist is delivered an FDA-cleared DICOM image (today) on the right, she is delivered an FDA-cleared draft report to edit (tomorrow) which will be the killer app? |
Radiology already provides a concrete example
In July 2026, FDA cleared the See-Mode Augmented Reporting Tool, Breast (SMART-B), marketed as DeepHealth Breast Ultrasound, under K260303. The FDA-hosted 510(k) summary explicitly describes generating a report from image-analysis results, followed by review and approval by a qualified physician, who can change the report before finalizing it. This is a specific breast-ultrasound application with structured reporting; it does not establish clearance of an unrestricted generative LLM. (This includes a 20-page FDA summary.) See [7,8].
The relevance to pathology is the workflow: image analysis feeds a draft report that a physician reviews and finalizes. (See figure above.) The FDA-funded Cognita project concerns how such outputs can be evaluated more comprehensively. Together, these developments make report generation a practical regulatory and commercial issue.
Why this could drive scanner installations
Consider a future prostate-biopsy application that organizes findings by specimen, proposes the supported diagnostic elements, and assembles the draft report. The pathologist would still examine the relevant slides, resolve uncertainty, order additional workup when needed, and take responsibility for the diagnosis. The commercial opportunity depends on making that complete workflow reliably faster and better.
An FDA-cleared radiology writer is said to save 40%, nearly half, of the radiologist's time.
That could change the purchasing decision for digital pathology. If report preparation saves meaningful professional time across a large routine caseload, the benefit can help justify the scanner, storage, interfaces, and software required to deliver it. Report automation could therefore drive hardware adoption as well as software revenue. A business case based on demonstrable productivity also need not depend entirely on securing a separate reimbursement payment for every AI step.
This remains my market thesis, not a result established by the FDA award. The critical measure will be the time and quality of the completed, physician-reviewed case, including corrections and exceptions.
Our RUO-LDT paradox may prolong the lag
Radiology’s head start is partly infrastructural: it already has widespread digital image workflows. Pathology must also digitize its inputs. But I suspect our own RUO LDT laboratory-developed-test tradition contributes another obstacle: LDT flexibility at the individual laboratory can produce a hard and fragmented route to widespread adoption.
- (That's why we read about 1200 FDA-approved radiology apps and 4 FDA approved pathology ones.)
Under CLIA, laboratories introducing unmodified FDA-cleared or approved test systems verify performance specifications; laboratories introducing noncleared systems or modifying cleared systems establish performance specifications. FDA clearance therefore does not eliminate local implementation work. It can, however, supply a common, manufacturer-supported evidence base and a defined intended use. The obligations for a particular AI application depend on its role in the testing workflow. [9]
The economic concern is repeated local effort, over and over, re-validating the same RUO. A sophisticated academic laboratory may have the cases, informatics staff, and expertise to validate a noncleared application. Many intermediate and smaller practices do not. If every possible customer must undertake a substantial validation project on the pathology RUO-LDT (but radiology uses FDA software), a technically impressive application may never become a practical purchase for much of the market. Pathology risks tying its own feet together: local flexibility can coexist with a very inefficient path to scale.
CMS and CDC are already examining the issue. Their July 16, 2026 CLIA request for information asks about AI in postanalytic interpretation, its role in interpreting histopathology slides, and methods for verifying software performance. That is a request for input, not a new AI validation rule, but it puts the implementation problem squarely on the policy agenda. [9]
What pathology should track
The immediate watch items are Cognita’s evaluation methods, radiologist-adjudicated error findings, and evidence of what large cohorts reveal that conventional validation studies miss. The larger question is whether these methods can help developers establish credible, transferable evidence for physician-supervised report generation.
For scanner manufacturers and pathology software companies, a validated report-generation application could become the reason a laboratory commits to digitization. FDA’s $1.29 million contract is a modest research investment, but it addresses a prerequisite for a much larger market: demonstrating that machine-prepared diagnostic reports can be evaluated rigorously enough for routine clinical use.
##
Sidebar; RVU valuation?
If AI saves the radiologist and pathologist time by 40%, will this cut MD work RVUs?
Don't hyperventilate. We have a case study for that. My academic (professorial) career spanned 1990-2001, which spanned at least the academic transition from "film" (2x3 foot films) reading of CT-MRI, to digital.
When did the RVUs change to catch up? About 2016 (when radiology tech RVUs went down a notch.) Here.
Sources
1. Quinn B. Is Digital Pathology Close to the Billion-Dollar Explosion in Value? September 1, 2026. Includes white-paper links. Read source
2. AuntMinnie. Cognita wins $1.29M FDA contract to evaluate radiology AI. September 16, 2026. Read source
3. Stempniak M. Rad Partners scores $1M FDA grant to test new way of evaluating AI-generated radiology reports. Radiology Business. September 17, 2026. Read source
4. Radiology Partners. Cognita Imaging Receives $1.29 Million FDA Contract to Test New Approach to Evaluating Generative AI in Radiology. September 16, 2026. Read source
5. FDA. Regulatory Science Extramural Research and Development Projects. Read source
6. Ostmeier S, et al. GREEN: Generative Radiology Report Evaluation and Error Notation. 2024. Read source
7. DeepHealth. DeepHealth Receives FDA Clearance for AI-Powered Breast Ultrasound. July 30, 2026. Read source
8. FDA. SMART-B clearance letter and 510(k) summary K260303. July 28, 2026. See PDF page 6 for report generation and physician review. Read source
9. CMS and CDC. Request for Information; Clinical Laboratory Improvement Amendments of 1988 (CLIA) Regulations. July 16, 2026. Sections on performance specifications and postanalytic interpretation. Read source
