HEADER:
Radiology AI is moving from narrow “finding detectors” to software that drafts full radiology reports for physician review. STAT’s coverage of Cognita, Aidoc, RSNA, and newly issued FDA breakthrough designations suggests a likely preview for digital pathology: AI-drafted prostate, GI, breast, skin, and biomarker reports are not here yet, but may be closer you thought.
See also Jerome Avondo at Linked In.
###
Chat GPT 5.5
###
Radiology AI Is Starting to Write Reports. Pathology Should Pay Attention.
For years, the familiar model for AI in medical imaging has been narrow: flag a pneumothorax to the top of queue, detect a nodule, count positive cells, measure tumor area, or highlight a suspicious region. These are useful tools, but they remain “spot detectors.” The physician still creates the clinical report.
Radiology is now moving toward something more consequential: AI systems that look at the image and draft the report.
That distinction matters. A narrow AI tool says, “possible pneumothorax.” A report-drafting AI might say, “small right apical pneumothorax, new since prior study,” and place that language into a preliminary report for a radiologist to review, edit, and sign. One is an alert. The other is an attempted clinical work product.
For digital pathology, this should be a wake-up call. If radiology is moving toward image-to-report generation, pathology may not be that far behind.
From Detection to Draft Reports
Three recent STAT articles by Katie Palmer describe a radiology AI field moving quickly from point solutions to foundation models and generative reporting. One article focuses on Cognita Imaging, a Stanford researcher-founded startup acquired for $80 million by the technology arm of Radiology Partners, the largest radiology practice in the United States. Radiology Partners includes more than 4,000 radiologists reading more than 55 million images a year, giving it unusual scale for training, testing, and deploying AI.
Cognita is not trying to build a single-finding detector. Its goal is broader: take in the current image, consider prior studies, identify important findings, describe location and severity, and draft report language. Within Radiology Partners, Cognita’s model has been incorporated into Mosaic Drafting, a tool that analyzes X-rays and head CTs and produces preliminary radiology report text for human review.
Cognita reports that a retrospective analysis of more than 95,000 exams showed about a 20% average reduction in radiologist time, with detection increases for some findings of up to 52%. The key claim is not “AI replaces radiologists.” It is “AI plus radiologist” may be faster and more accurate than radiologist alone.
That is exactly the claim digital pathology would like to make. Not autonomous sign-out. Not black-box diagnosis. Rather: AI drafts routine report elements, highlights the evidence, quantifies key features, and lets the pathologist become editor, verifier, and final signer.
The RSNA Lesson:
AI Is Everywhere, but Deployment Is Hard
A second STAT article, reporting from the 2025 Radiological Society of North America meeting, describes AI as nearly unavoidable in radiology. More than 100 companies filled the AI showcase. Major imaging companies are embedding AI into hardware and workflow. Vendors are pitching tools for fracture detection, mammography, brain atrophy, coronary imaging, emergency triage, report summarization, and report generation.
Why? Imaging volume is rising, radiologists are scarce, and health systems need more capacity.
But the article is also cautionary. Even older, narrow FDA-cleared radiology AI tools often require significant local work before they perform well. FDA clearance does not mean plug-and-play success. Algorithms must be integrated into PACS, reporting systems, worklists, and physician behavior. They must be monitored. False positives can create alert fatigue.
Digital pathology should absorb this lesson early. A pathology AI module that suceeds in a peer review paper is not the same as a pathology AI system that improves daily sign-out. The practical questions are decisive: Does it fit into the LIS? Does it work in the viewer? Does it slow the pathologist down? Does it create extra clicks? Does it improve the report, or does it create a second job checking the machine?
Radiology is shifting from isolated algorithms to platforms: systems that manage deployment, integration, validation, monitoring, workflow, and reporting. Digital pathology is likely to follow the same path. The winning product may not be one algorithm for one stain. It may be an enterprise layer that connects images, AI outputs, structured fields, report text, QA, audit trails, and regulatory monitoring.
FDA Is Beginning to Engage, Awards BT Status
The third STAT article reports that FDA has granted breakthrough designation to two generative AI radiology devices that interpret chest X-rays and draft reports: one from Cognita and one from Aidoc. Breakthrough designation is not FDA authorization, but it matters because it shows FDA is now formally engaging with generative image interpretation as a medical device category.
This is a major shift. FDA is more comfortable with structured outputs: yes/no, lesion present/absent, measurement above or below threshold. Report drafting is much more complex. A report can contain many findings, modifiers, measurements, comparisons, exclusions, and clinical judgments. It can also contain hallucinations, omissions, and subtle wording errors.
That is why validation is hard. A narrow device can be tested for whether it detects pulmonary embolism or fracture. A draft report must be evaluated for whether it is clinically acceptable as a whole. Did it miss an important finding? Did it invent one? Did it locate the lesion correctly? Did it compare to prior imaging correctly? Did it overstate certainty?
Ordinary automated natural-language-processing metrics do not solve this problem. The “ground truth” human report may itself contain errors. Two reports may use different words but mean the same thing. Conversely, two reports may look almost identical but differ in one decisive word, such as “no” before “metastasis,” “perineural invasion,” or “carcinoma.”
The validation problem is therefore not merely computational. It is medical, linguistic, regulatory, and workflow-based.
The Immediate Pathology Translation
A pathologist reading these articles will immediately imagine the surgical pathology version.
For prostate biopsy:
“Prostatic adenocarcinoma, Gleason score 3+4=7, Grade Group 2, involving 70% of one core; pattern 4 comprises approximately 10% of tumor. Perineural invasion not identified.”
For colon biopsy:
“Tubular adenoma with low-grade dysplasia. No high-grade dysplasia or invasive carcinoma identified.”
For breast core biopsy:
“Invasive ductal carcinoma, Nottingham grade 2, greatest linear extent 8 mm in submitted cores; associated intermediate-grade ductal carcinoma in situ present.”
This is not science fiction. Digital pathology AI already performs pieces of this work: tumor detection, gland segmentation, biomarker quantification, suspicious-region highlighting, tumor measurement, and selected grading tasks. The next step is to connect those observations into draft diagnostic language.
Prostate biopsies are an especially vivid use case. The reports are repetitive, structured, and quantitative. A pathologist records site, carcinoma present or absent, Gleason patterns, Grade Group, number of involved cores, percent involvement, tumor length, perineural invasion, cribriform or intraductal carcinoma, high-grade PIN, ASAP, and comments. Many of those elements are exactly the kind of structured observations an AI system could propose.
The killer app may not be “AI diagnoses prostate cancer.” It may be: “AI pre-populates the prostate biopsy report, including grade suggestion, tumor length, percent involvement, and core-level mapping, for pathologist review.”
Why Pathology May Be Easier — or Harder
Pathology has some advantages. Many pathology reports are already semi-structured. CAP synoptic templates, biomarker templates, AJCC fields, and repeated diagnostic phrases create a controlled reporting grammar. A pathology report-drafting AI does not need to write elegant prose. It needs to populate the right fields and use conventional diagnostic language.
Pathology also has many high-volume, repetitive specimens: prostate cores, GI biopsies, skin shaves, breast cores, bladder biopsies, cervical biopsies, and lymph nodes. These are natural targets for draft tempates, not entertaining prose.
But pathology also has unique obstacles. Preanalytic variability is enormous: fixation, processing, embedding, section thickness, stain intensity, scanner type, focus quality, tissue folds, crush artifact, cautery artifact, decalcification, and lab-specific staining culture. A model trained on one laboratory’s H&E slides may behave differently on another laboratory’s slides.
Pathology diagnosis is also deeply contextual. The same microscopic pattern may mean different things depending on specimen site, clinical history, prior diagnosis, immunostains, molecular results, gross description, and sampling. Some use cases are bounded; many are not.
Then there is automation bias. If an AI draft says “Grade Group 1,” will the pathologist be less likely to notice a small focus of pattern 4? If it says “no perineural invasion,” how carefully will the pathologist search? It seems plausible that AI text can change how humans read.
The Hard Part Comes After the Demo
Radiology’s experience suggests that generating a plausible report is not the hardest part. The harder question is whether the draft report is safe, useful, and better than current workflow.
For pathology, validation will need to test image-level detection, quantification, classification, report language, scanner variability, stain variability, inter-laboratory performance, and human behavior. It will also need post-deployment monitoring, because performance may drift as stains, scanners, case mix, and users change.
FDA clearance or authorization will not answer all of those questions. Local validation and real-world monitoring will still matter.
The Business Case
Radiology also teaches a payment lesson. Many radiology AI tools do not generate separate reimbursement. They must, and can, justify themselves through productivity, quality, triage, turnaround time, or reduced downstream cost.
The same will likely be true in pathology. AI report drafting may not receive a separate CPT code or per-case Medicare payment. It will be purchased if it lets a pathology group handle more cases, reduce turnaround time, standardize reporting, improve quantification, reduce addenda, or support subspecialty quality assurance.
A mediocre draft that requires constant correction will not save time. A highly reliable draft for selected high-volume specimens might.
Bottom Line
Radiology is moving from AI that flags findings to AI that drafts reports. FDA has begun engaging with this category. Large radiology practices are already testing such tools under research protocols. Vendors are racing toward foundation models that can turn images into structured clinical language.
Digital pathology should view this as a preview.
The near-term future is probably not autonomous pathology sign-out. It is AI-generated draft reporting for bounded, high-volume use cases: prostate biopsies, GI biopsies, breast cores, skin specimens, lymph nodes, and selected biomarker workflows. The pathologist remains responsible, but the starting point changes. Instead of a blank report, the pathologist receives a structured draft: diagnosis, measurements, grade, extent, and key negatives, ready for review and correction.
The challenge is not imagining this future. It is making it safe, useful, trusted, and economically compelling.
$idebar: Could AI-Drafted Reports Eventually Reduce RVUs?
Radiologists may see an RVU issue on the horizon. If AI-drafted reports become reliable enough to finalize more quickly than manually dictated reports, CMS or the RUC could eventually ask whether the physician work RVU should be lower. (This happened with RVUs after the industry had slowly but almost completely transitioned to net-cost-saving PACs).
That concern is not fanciful. Physician work RVUs reflect time, intensity, complexity, and professional effort. Radiology has been through a version of this before: the shift from wet-film radiology to digital PACS was associated with a downward RVU adjustment, particularly focused on wet-film costs. A future reviewer could make a similar argument if radiologists are primarily editing AI-drafted reports rather than creating reports from scratch.
However, there is no clear sign of that yet. In adjacent clinical documentation settings, early experience with ambient AI for E&M visits has not shown any significant net reduction in clinician time to finalize notes. The clinician still has to review, correct, personalize, and take responsibility for the final record.
The same may prove true in radiology, and later in pathology. A partly reliable draft can create new work: checking the AI, verifying measurements, correcting wording, and guarding against automation bias. The RVU impact will depend not on whether AI drafting exists, but on whether it measurably and durably reduces physician work in real-world practice.
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/