Weekly AI digest
Radiology & medical imaging AI · week of 27 September to 3 October 2026
207 peer-reviewed papers · 5 industry and regulatory items · conference and KOL highlights. Adapted from my weekly intelligence report. Full report: FR (PDF) · EN (PDF) · RU (PDF).
Governance takeaway
What to do about it this week
Stop evaluating AI tools on average performance and start evaluating them on where the average breaks: this week supplies the instruments to do it. Put two new lines in your pre-deployment template, performance by subgroup (with missing subgroup evidence logged as a risk, not a footnote) and end-to-end latency against the report-finalization window per modality.
The JMIR refined-exclusion gates and the NHS 8-domain checklist give you the structure; the 20-center latency study and the breast AI agreement data show exactly what those lines will catch.
Regulation & AI Act watch
Visiana's BoneXpert cleared for the US
FDA 510(k) clearance (K262390) to support pediatric radiologists in assessing bone age from hand radiographs, ages 2 to 21, Greulich-Pyle method; CE-marked and in European use since 2009. The question to ask: test-retest consistency figures, because the clinical question is usually how bone age changes between visits, not a single reading.DeepHealth clears a foundation-model chest X-ray CADe
510(k) for Chest XRay, detecting and localizing suspected thoracic abnormalities on radiographs, one of the first US-cleared chest CADe tools built on a foundation model; the evidence cited publicly is a company release plus European deployment data. The question to ask: what the Predetermined Change Control Plan commits to when the underlying foundation model is updated.CapsoVision's AI Highlights cleared on a competitor's predicate
510(k) for an AI reading aid that flags suspected lesions in small-bowel capsule endoscopy for suspected bleeding; the predicate is a rival's De Novo-authorized device, and no performance data have been published. The question to ask: the reader-study endpoints and the miss rate when physicians review only highlighted frames.Stockholm Region replaces the second reader with AI
Lunit INSIGHT MMG becomes the independent second reader for 200,000 to 250,000 annual breast screening exams across three Stockholm hospitals, on the strength of the prospective ScreenTrustCAD study and three years of routine use. The question to ask: what the monitoring plan tracks once the human second reader is gone, and who can suspend the AI reader.Clairity Breast goes direct-to-consumer
Everlywell now sells nationwide access to Clairity Breast, the FDA De Novo-authorized tool estimating five-year breast cancer risk from a screening mammogram; De Novo means no predicate existed. The question to ask: how a consumer-ordered risk score re-enters the clinical record and who absorbs the downstream supplemental-screening demand.
Evidence you can use
Selected from 207 papers indexed in PubMed this week.
Reducing Clinician Annotation Fatigue in Open-Set Federated Learning: A Client-Adaptive Vision-Language Gatekeeper
Journal of the American College of Radiology, 01 October 2026Deep Learning Analysis of Dual-Modality US Videos for the Characterization of Superficial Lymphadenopathy
Radiology: Artificial Intelligence, 30 September 2026AI Latency, Report Turnaround Time, and Adoption in a Multi-Vendor AI Ecosystem: A Multi-Site Observational Study
Journal of the American College of Radiology, 29 September 2026Tomosynthesis and Synthesized-2D Breast AI Outputs in a Biopsy-Referred Cohort: A Diagnostic Study of Agreement and Incremental Decision-Support Value
Journal of Imaging Informatics in Medicine, 29 September 2026Refined Exclusion in Medical AI: Data Justice and Patient Safety Governance
Journal of Medical Internet Research, 01 October 2026Dual-discriminator Generative Adversarial Network (DD-GAN) for Virtual Elastography Ultrasound in Thyroid Cancer Diagnosis
Academic Radiology, 28 September 2026Evaluating the Long-Term Malignancy Risk Prediction Accuracy of Major Guidelines for Pancreatic Cystic Lesions Using Radiology Features
Journal of the American College of Radiology, 28 September 2026Translating Real-World Safety and Implementation Gaps Into a Deployment-Derived AI Readiness Preimplementation Checklist for NHS Health Care Providers
JMIR AI, 28 September 2026
Post-market signal
Legal analyses published this week by Cooley (28 September) and Haynes Boone (2 October) converge on the significance of the FDA CDRH discussion paper on generative-AI-enabled devices: the agency is openly considering accepting greater premarket uncertainty in exchange for heavier reliance on post-market monitoring, through periodic benchmarking against premarket thresholds, sample-based independent clinician review of real-world inputs and outputs, and performance-degradation monitoring. CDRH even asks whether machine-based supervisory agents could carry part of that oversight.
Early commenters are pushing back that post-market monitoring is no substitute for premarket evidence where harm is severe or irreversible, and want sponsors to state which failure modes a monitoring program cannot detect. Every proposed mechanism presumes the hospital can supply or host the monitoring data, which makes this a deployer issue, not just a vendor one. Comments close 19 October 2026 (docket FDA-2026-N-7874); imaging departments with deployed AI have standing to comment and rarely do.
Playbook snippet
One step for your AI committee
Anchor: EU AI Act Article 72 (post-market monitoring); ISO/IEC 42001 performance evaluation clause. The AI Act application calendar is under revision following the Digital Omnibus proposal, so verify dates against the current text.
Step: Sample 20 AI outputs per month against the final signed report. Track one number: concordance rate. Plot it. Add a second line this month: the share of AI results that arrived after report finalization, per modality. A trend, not a threshold, is what tells you something has changed.
Why: Drift is invisible in aggregate accuracy and visible in a monthly line. Starting the series now means you have history when the obligation bites; starting later means you have a snapshot.
Worked example: Morozov S et al. (Journal of the American College of Radiology, PMID 42810607, DOI 10.1016/j.jacr.2026.09.026) show what the second line catches: across 96,874 examinations, 7.2% of AI results arrived after the report was signed, from 3.0% for knee MRI to 13.2% for chest CT. A tool whose output misses the reporting window is drifting operationally even when its accuracy has not moved.
From the field
The ACR-SIIM Practice Parameter for Imaging Artificial Intelligence entered into effect on 1 October under the ACR's standard annual cycle, making it the reference US practice standard from this week: AI governance group, tool inventory with versions and intended use, local acceptance testing, real-world performance monitoring with stop rules, and explicit coverage of non-FDA-regulated tools including generative AI. RSNA meanwhile shipped version 2.0 of its Foundational AI Certificate program, a structured and auditable way to evidence workforce AI literacy.
Nina Kottler answered the question nobody dares ask out loud, what a Chief Medical AI Officer actually does, with a two-month field report: site visits, teaching, and writing on how AI changes clinical decision-making, how to implement it responsibly and how to monitor it in use, a concrete sketch for any organization weighing whether to create the role.
Woojin Kim turned his 2024 Radiology editorial into a 25-second animated explainer with a generative model and shared the prompt, along with the caveat that matters: the model confidently mislabelled an absolute-washout value, and seeing what is wrong still takes a human.
Amine Korchi took the three-scenarios framework for radiology's AI future, expansion, concentration and stratification, to a teaching session at Hospital Israelita Albert Einstein in São Paulo, a sign of how actively radiology leadership audiences worldwide are working through AI scenario planning.
Source articles are indexed in PubMed with verified DOIs. Manuscript-stage work is excluded. The full weekly report is produced in French.
Get the weekly digest
One email a week on what matters in radiology and imaging AI. Researched by a radiologist and imaging informatics professional. Free, no vendor sponsorship.