← All issues

Weekly AI digest

Radiology & medical imaging AI · week of 13 to 19 September 2026

202 peer-reviewed papers · 2 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

This week makes one deployment lesson concrete: neither clearance nor good aggregate accuracy decides where a tool should run. A prospective shadow-mode study of an FDA-cleared aneurysm detector found operational value strongly favorable in inpatients, marginal in the emergency department and unfavorable in outpatients, with the same algorithm in the same hospital. An expert panel on AI ARIA detection reached the matching policy answer from the governance side: conditional implementation only, with a radiologist in the loop, quality assurance and prospective performance monitoring.

The action for a radiology leader: write deployment scope per care setting into acceptance tests, and use shadow mode to measure incremental yield and gain-to-pain in each setting before the tool touches a live worklist.

Regulation & AI Act watch

Evidence you can use

Selected from 202 papers indexed in PubMed this week.

Post-market signal

No new post-market surveillance event (drift report, recall, MAUDE signal or Article 72 guidance) surfaced in the window, so this section is labelled commentary on this week's evidence rather than reported news.

Two of this week's papers show why a single monitoring line per tool is not enough. The JACR shadow-mode study found the same cleared aneurysm algorithm operationally favorable in inpatients (gain-to-pain 2.57) and unfavorable in outpatients (0.67); a pooled concordance metric would have averaged those into a false reassurance. The FAR-POLYP-SEG benchmark found false-positive rates on normal tissue inverting the internal accuracy ranking of six segmentation models. The monitoring consequence: stratify your monthly concordance series by care setting, and track false positives on studies the final report calls normal as their own line. Aggregate concordance can hold steady while one setting quietly degrades.

Playbook snippet

One step for your AI committee

Anchor: MDR and FDA intended-use statements; conformity assessment.

Step: For every cleared tool you run or plan to buy, copy the intended-use sentence verbatim to the top of its deployment file. Under it, list every way your intended use differs: population, modality, scanner fleet, care setting, reader mix. Treat each listed difference as one named local validation task with an owner and a date.

Why: Clearance is a statement about a specific claim on specific evidence, not about your deployment. A viewer clearance is not a detection claim, and a De Novo authorization means no predicate device exists, so the local validation burden is higher, not lower.

Worked example: Goldberg-Stein S et al. show what one such difference is worth: the same FDA-cleared aneurysm algorithm, measured prospectively in shadow mode over 3,856 CTAs, returned a gain-to-pain ratio of 2.57 in inpatients and 0.67 in outpatients. The care-setting line in the deployment file flips the decision by itself. JACR, PMID 42742498, DOI 10.1016/j.jacr.2026.07.018.

From the field

The week's institutional news came from outside radiology's usual venues. HHS named the first radiologist ever to the US Preventive Services Task Force on 17 September, diagnostic radiologist Dennis Wulfeck, and the ACR welcomed an appointment it has long advocated for; task force composition shapes which imaging-based screening services get graded and covered. A day earlier, Commission President von der Leyen made AI-supported mammography the lead health example in her State of the Union address, and the ESR responded in kind.

EuSoMII held a live "CBR meets EuSoMII" joint session on 18 September at the 55th Brazilian Congress of Radiology in Recife, extending European imaging-informatics collaboration into Latin America ahead of its 2026 annual meeting in Crete.

On the expert side, Curt Langlotz amplified Stanford AIMI's MedVAL paper in npj Digital Medicine, a framework for expert-level validation of AI-generated medical text, with the pointed observation that generation is now fast and verification is the bottleneck. Geneva neuroradiologist Amine Korchi highlighted a new comparative evaluation in Chest of generative models drafting emergency chest X-ray reports, showing confident but divergent narratives from the same image.

Trade media closed the loop on the week's lead paper: AuntMinnie's coverage of the JACR aneurysm shadow-mode study reached its most-read list, and Diagnostic Imaging's Reading Room podcast devoted its second episode on AI reimbursement to why the business case for imaging AI still has to be measured locally.

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.