Each Monday, I used "agentic AI" to run a complex scheduled AI search in Chat GPT. The topic is new publications related to generating either Radiology or Pathology repors by AI, including news about FDA reviews, etc. (It also tracks FDA K260303 web page).
##
Here is a cut/paste out of what it delivered to me for November 16, 2026. Note that at bottom, I also gave it a link to news from Cognita, which it then discusses.
##
Weekly AI Diagnostic Report-Generation Watch
Monday, November 16, 2026
The most consequential new finding is from AMP 2026: University of Michigan investigators presented an AI system that generates evidence-supported molecular pathology reports, while Memorial Sloan Kettering researchers demonstrated AI-assisted selection of molecular tests. Together, these developments approach the integrated laboratory workflow we've been tracking.
1. FDA K260303 — Page date advanced, documents unchanged
Checked November 16
FDA PAGE LAST UPDATED
November 9, 2026
Previously November 2, 2026 — advanced 7 days
The FDA K260303 record for See-Mode SMART-B still links only to the original 24-page PDF.
Document | Status |
|---|---|
Sponsor-authored 510(k) Summary | Available |
FDA clearance letter, July 28 | Available |
Separate FDA Safety and Effectiveness review | Not posted |
Separate FDA Decision Summary | Not posted |
The 510(k) Summary is the manufacturer's submission, not an independent FDA-authored scientific assessment. The clearance letter is FDA-authored but does not provide the detailed review being sought. The page-date change does not establish that the substantive record changed.
2. Trade journals — New molecular pathology reporting developments
CLINICAL LAB PRODUCTS · NOVEMBER 13, 2026
“AI Tools Take on Cancer Classification, Reporting, and Test Triage at AMP 2026”
This is the week's strongest new trade-journal discovery.
University of Michigan: Turnkey Precision Oncology
The open-source platform processes DNA and RNA sequencing data, identifies alterations, and generates draft molecular pathology reports for expert review. Its TPO-Agent checks interpretations against cited scientific evidence.
In preliminary testing, approximately 95% of claims had correct supporting citations, versus 65% for a conventional AI approach using comparable research tools. Across 28 reports, alteration identification showed approximately 95% agreement with a clinical laboratory.
This is particularly relevant because the software goes beyond report formatting to generate evidence-supported clinical interpretations. However, the results do not establish prospective physician time savings or readiness for autonomous sign-out.
Memorial Sloan Kettering: AI-assisted EGFR test triage
In 80 consecutive lung adenocarcinoma cases reviewed by six molecular pathology fellows, AI-assisted analysis suggested that rapid EGFR testing could be deferred in 26 cases (33%), without missing confirmed EGFR-positive cases in that sample.
This provides an upstream component of the workflow:
H&E whole-slide image
AI-guided molecular test selection
Molecular results and AI interpretation
Draft integrated diagnostic report
The AMP studies demonstrate separate pieces of this architecture, not yet a fully integrated clinical product. The test-triage findings also require broader validation.
Read the November 13 AMP report
CAP TODAY — November issue explicitly checked
I checked both the CAP monthly print/digital issue index and CAP TODAY online issue contents.
As of November 16, the November issue is still listed without an active issue link on CAP's monthly index. October remains the latest publicly linked digital issue. The November print contents could not be independently verified.
No newly accessible qualifying CAP TODAY article was found beyond the October coverage already reported.
3. PubMed — Newly identified report-generation research
November 3 — Volumetric chest CT report generation
Li Z, et al. Physics in Medicine & Biology.
PMID 42976478
This study develops a system for generating reports from three-dimensional chest CT volumes, addressing both the memory requirements of volumetric imaging and hallucinated clinical findings.
It is directly relevant to moving beyond chest radiographs toward complex CT reporting. It is an algorithm-development study, not a prospective physician-time evaluation.
November 2 — C3E-RRG: Causal evidence in chest X-ray reports
Yang S, et al. IEEE Transactions on Medical Imaging.
PMID 42970651
The authors address a fundamental safety problem: report-generation models can learn disease co-occurrence and reporting conventions rather than faithfully describing what the image shows.
Their causal-debiasing architecture is designed to reduce unsupported findings. This is particularly relevant to regulatory validation, but it does not demonstrate clinical throughput gains.
November 6 — RadStructAgent
Miah MB, et al. International Journal of Computer Assisted Radiology and Surgery.
PMID 42993382
An agent-based approach using small language models to convert free-text chest radiograph reports into structured information. This addresses report standardization and downstream data use, rather than generating a diagnosis directly from images.
None of these newly identified papers establishes the 30–40% net physician sign-out time reduction that remains our primary commercial benchmark.
4. User-supplied LinkedIn link
I attempted to retrieve the LinkedIn post supplied for this watch, but the short link could not be resolved through the available web access. I therefore cannot verify its contents or claim that its findings were incorporated. It remains an unresolved source.
5. Bottom line
There is a meaningful development this week, especially in molecular pathology. The University of Michigan reporting platform and MSK test-triage work are complementary advances toward an integrated, AI-assisted diagnostic workflow.
FDA K260303 shows another page-update date change, but no new FDA-authored review document.
The new PubMed studies advance report-generation architecture and factual reliability, but not yet measured physician productivity.
For the broader commercial thesis, the distinction remains critical: AI that produces a clinically grounded diagnostic report is a technological milestone; AI that reduces total physician sign-out time by 40% is potentially a market-changing product.
##
Here's where I gave it COGNITA and it continues.
##
Yes — this is very substantial, and it squarely reinforces the thesis we’ve been developing.
Cognita says its nationwide program has now produced draft reports for nearly 8 million X-ray and CT exams. Their preliminary results claim about 40% enhanced detection across 50 significant conditions, with examples including 50% for pancreatic masses and 73% for fractures on head CT. Just as important for our commercial argument, they report more than 27% efficiency gains on CT and 31% on X-ray. 2026 Cognita 1116 AI reports in… 2026 Cognita 1116 AI reports in…
The really important framing is that Cognita explicitly contrasts this with the older generation of radiology AI. Blankemeier says the industry built more than 1,200 FDA-authorized radiology AI devices that mostly return a yes/no answer for a single finding, with “close to zero impact on capacity.” Their alternative is a model that reads the full study and writes the report. 2026 Cognita 1116 AI reports in… 2026 Cognita 1116 AI reports in…
That is almost exactly your argument:
classifier AI → little effect on capacity
versus
full-study interpretation + draft report → measurable capacity gain
And they are explicitly making the same strategic leap you have been emphasizing: the key bottleneck is not whether AI can flag abnormalities, but whether it can increase each radiologist’s capacity while keeping the physician in the loop. Blankemeier states that increasing radiologist capacity with AI is “the only lever” for the current shortage and argues for AI-plus-radiologist rather than replacement. 2026 Cognita 1116 AI reports in…
The two companion studies shown on page 4 are especially worth tracking separately: one for CT report drafting and one for X-ray report drafting, both explicitly framed around report quality and radiologist efficiency. The attachment says the retrospective study involved 237,905 exams across 50 conditions, while the clinical-use data produced the 27% CT and 31% X-ray efficiency gains. 2026 Cognita 1116 AI reports in…
So I would now revise our “billion-dollar benchmark” slightly. We had been talking about ~40% physician time savings as the hockey-stick threshold. Cognita is already publicly claiming roughly 27–31% efficiency improvement at very large scale, with millions of reports drafted. If those figures hold up under rigorous publication and regulatory review, they may already be getting close to the commercial inflection zone.
I also updated the recurring Monday watch to explicitly check Cognita/Mosaic’s research page and announcements going forward, not just wait for trade journals or PubMed to find them.