Radiology AI Is Starting to Write Reports. Pathology Should Pay Attention.
[Claud Opus 4.8]
Radiology AI is crossing the line from flagging findings to drafting the report. STAT's recent coverage of Cognita, Aidoc, RSNA 2025, and two new FDA breakthrough designations is, for digital pathology, a preview of the next three years — including the awkward part nobody markets: a draft report doesn't save the physician's time so much as it relocates that time into editing.
For a decade, AI in medical imaging has meant a spot detector. Flag the pneumothorax. Count the positive cells. Measure the tumor. Highlight the suspicious region. Useful, narrow, and bounded — and in every case the physician still writes the report.
That line is now being crossed. A narrow tool says "possible pneumothorax." A report-drafting model says "small right apical pneumothorax, new since prior study," and drops that language into a preliminary report for a radiologist to review, edit, and sign. The first is an alert. The second is an attempted clinical work product. The difference is the whole story.
Three STAT articles by Katie Palmer trace the move. Worth reading as a set, because they form an arc that pathology is likely to walk in turn: a marquee acquisition, a field-wide reality check, and the FDA beginning to engage.
What radiology is actually doing
The acquisition is Cognita Imaging — a Stanford researcher-founded startup bought late in 2025 for $80 million by the technology arm (Mosaic Clinical Technologies) of Radiology Partners, the largest radiology practice in the country. The scale is the point. Radiology Partners fields more than 4,000 radiologists reading more than 55 million images a year, and serves roughly 12% of the U.S. market against a second-place group at about 2%. For training a vision-language model on images paired with their written reports, that asymmetry is the moat, and CEO Louis Blankemeier says so plainly: comparable data exists, but not in one place.
Cognita's model is embedded in a tool called Mosaic Drafting, which reads X-rays and head CTs and produces preliminary report text. The headline claims, from a retrospective look at more than 95,000 exams: about a 20% average reduction in radiologist time and detection increases on selected findings of up to 52%. The framing is deliberately modest — not "AI replaces the radiologist" but "AI plus radiologist," faster and more sensitive than the radiologist alone. Deployment has been staged and run under research IRB: chest X-ray first, then the rest of plain film, abdomen, MSK, and most recently head CT.
The second article, from the RSNA 2025 conference, supplies the reality check. More than 100 companies packed the AI showcase; the hardware majors (GE, Philips, Siemens) have leaned in; nearly every booth now calls itself a "platform." But the meeting rooms told a quieter story than the demo floor. Even older, narrow, FDA-cleared tools routinely need substantial local work before they perform — Stanford's David Larson called it "some assembly required." More algorithms mean more false positives, more alert fatigue, and more integration pain across PACS, worklists, and reporting systems. in radiology, almost none of it bills: with few AI tools generating separate reimbursement, the purchase has to pay for itself in time, turnaround, or downstream cost — at which point, as Emory's Hari Trivedi noted, it can simply be cheaper to hire another radiologist.
[BQ In the shift from wet-film to digital CT-MRI, that didn't bill either, but was driven by appreciable savings, as shown by a 5% valuation drop when AMA RVUs shifted from wet-film practice expense RVUs to digital PACs RVUs.]
The third article, from June 2026, is the regulatory turn. The FDA granted breakthrough designation to two generative draft-reporting devices — Cognita in March, and Aidoc's "First Read" (scoped to four life-threatening findings) in June. Breakthrough status is not authorization; it is a standing line to the agency while the rules get written. But it signals that the FDA now treats generative image-to-report as a device category, not a curiosity. Cognita, Aidoc, Voio, and Harrison.ai are all pushing toward CT-scale foundation models, and all of it is running today only under IRB research protocols.
The part the demo doesn't show
Here is the through-line pathology should internalize, because it is the one the marketing voices don't mention.
Generating a plausible report turns out not to be the hard part. The hard part is everything after: is the draft acceptable as a whole? A narrow device is validated on one axis — does it catch the embolism, yes or no. A draft report has to be evaluated for missed findings, invented ones, mislocated lesions, botched comparisons to priors, and overstated certainty, all at once. Stanford's Amy Hong, who has assessed more than ten of these models, found them weak at measuring and locating findings, still capable of hallucination and bias, and possessed of distinctly different "personalities" depending on whether their training data came from outpatient centers, ERs, or ICUs. Aidoc's Elad Walach put the economic version bluntly: the accuracy bar for usable augmentation is high. A draft you don't fully trust buys you nothing, because you re-check everything.
And this is where the E&M experience should temper everyone's enthusiasm. Ambient AI scribes for office visits were supposed to give clinicians their evenings back. In practice, the early evidence shows little net reduction in time to finalize the note. The clinician still reviews, corrects, personalizes, and — crucially — signs and owns the record. The physician's labor didn't disappear. It changed shape, from composition to editing and quality assurance. A 20% retrospective time estimate is a hypothesis about a workflow, not a law of nature, and a partly-reliable draft can manufacture its own new work: verifying measurements, hunting for the omission, and guarding against the temptation to just click "accept."
The pathology translation
A pathologist reading these articles will immediately picture the surgical-pathology version, and will be right to. The targets are obvious: prostate cores, GI biopsies, skin shaves, breast cores, bladder and cervical biopsies, lymph nodes — high-volume, repetitive, and already semi-structured. CAP synoptics, biomarker templates, and AJCC fields have already built a controlled reporting grammar. A drafting model in pathology doesn't need elegant prose; it needs to populate the right fields with conventional diagnostic language.
The prostate biopsy is the good case. The report is a near-checklist — site, carcinoma present or absent, Gleason patterns, Grade Group, cores involved, percent involvement, tumor length, perineural invasion, cribriform/intraductal, high-grade PIN, ASAP. This is structured writing, not creative writing. Digital pathology AI already can do fragments of this: tumor detection, gland segmentation, biomarker quantification, region highlighting, measurement. The plausible "killer app" is not AI diagnoses prostate cancer. It is AI pre-populates the prostate report — grade suggestion, length, percent, core-level map — for pathologist review.
But pathology also carries obstacles radiology doesn't. Pre-analytic variability is enormous and local: fixation, processing, section thickness, stain intensity, scanner, focus, folds, crush and cautery artifact, decalcification, and each lab's idiosyncratic staining culture. A model tuned on one laboratory's H&E may misbehave on another's. Diagnosis is also deeply contextual — the same pattern means different things by site, history, prior diagnosis, immunostains, and molecular results. And the automation-bias problem is sharper when the draft asserts a negative. If the model writes "Grade Group 1," does the pathologist still seek out the small focus of pattern 4? If it writes "no perineural invasion," how hard and for how long do you look? One decisive word — a "no" in front of "metastasis," "invasion," or "carcinoma" — is the entire diagnosis, and ordinary language-overlap metrics are blind to that.
So pathology's validation burden is the radiology burden plus a layer: detection, quantification, classification, and report language, across scanner and stain variability, across laboratories, with human-behavior testing and post-deployment monitoring for drift as scanners, stains, case mix, and users change. FDA authorization will not retire those questions. Local validation and real-world monitoring will.
The business case — and the RVU shadow
Expect the radiology payment lesson to repeat. AI report drafting will probably not earn a separate CPT code or per-case Medicare payment. It will be bought, if at all, because it lets a group handle more cases, shorten turnaround, standardize reporting, cut addenda, or backstop subspecialty QA. A mediocre draft that needs constant correction fails that test on contact. A highly reliable draft for a few bounded, high-volume specimens might pass it.
Sidebar — the RVU question.
If AI-drafted reports become reliable enough to finalize faster than dictation, CMS or the RUC could eventually ask whether the physician work RVU should fall. The precedent exists: the wet-film-to-PACS transition carried a downward adjustment tied to film-era costs. But there is no sign of it yet, and the E&M analogy cuts against it — if editing an AI draft consumes as much physician time as creating the report did, the work component hasn't actually shrunk. The RVU exposure turns not on whether AI drafting exists, but on whether it measurably and durably reduces physician work in the real world. So far, that case is unproven.
Bottom line
Radiology has moved from AI that flags to AI that drafts; the FDA is engaging; the largest practices are running it under research protocols. Digital pathology should read this as its own preview. The near-term future is not autonomous sign-out — it is structured draft reporting for bounded, high-volume specimens, with the pathologist still responsible but starting from a populated template instead of a blank one.
The hard part is not imagining that future. It is making it safe, useful, trusted, and economically compelling — and being honest that, on current evidence, the first thing a good draft changes is not how much the physician works, but what the physician's work is.
Sources
STAT Plus — Katie Palmer, "Off an $80 million acquisition, Cognita's CEO on the power of scale in radiology AI," Dec. 1, 2025. https://www.statnews.com/2025/12/01/cognita-imaging-radiology-partners-what-next-vision-language-models/
STAT Plus — Katie Palmer, "In radiology, an early adopter of AI, technology is advancing faster than the field can keep up," Dec. 5, 2025. https://www.statnews.com/2025/12/05/radiology-artificial-intelligence-advancing-faster-rsna/
STAT Plus — Katie Palmer, "FDA gives generative AI in radiology two breakthrough designation nods," June 25, 2026. https://www.statnews.com/2026/06/25/radiology-generative-ai-cognita-aidoc-fda-breakthrough-designation/
