← All issues

Weekly AI digest

Radiology & medical imaging AI · week of 23 to 29 August 2026

198 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

This week's evidence cluster is operational AI, the tools that never reach your AI committee because nobody calls them clinical: a no-show model that decides which patients get outreach along a deprivation gradient, a camera system screening MR-unsafe wheelchairs, an LLM routing extraspinal findings to specialists. Pick one operational tool that is live or planned in your department this month and write its acceptance test around the harmful error direction: the high-risk patient not flagged, the unsafe chair called safe, the finding downgraded out of the follow-up queue.

The wheelchair study shows the method: report the harmful direction on its own denominator (8.3% of MR-unsafe instances missed in the worst-case model, invisible inside a 0.702 mAP), then decide.

Regulation & AI Act watch

Evidence you can use

Selected from 198 papers indexed in PubMed this week.

Post-market signal

Two developments this week, both about who catches AI errors after go-live. Patient safety organization ECRI expanded its Problem Reporting Network, the confidential device-problem channel it has run since 1972, to explicitly capture errors, malfunctions and near misses involving AI tools, and called on providers nationwide to submit incidents. Its survey of 124 quality and safety leaders explains why: 31% encountered an AI output they believed incorrect in the past year, 9% said an error reached a patient or care decision, and 35% did not know. Alongside MAUDE and vendor complaint files, there is now a neutral place to send the near miss your radiologist caught.

Separately, the FDA posted a Class II recall record for Fujifilm Synapse PACS version 7.4.310: a software design flaw can make ruler measurements inaccurate on the reading workstation, with the workaround pointing at M-mode ultrasound measurements. Only six units are affected, but the lesson does not scale with the unit count: measurement is now a software function, and a measurement defect does not announce itself the way a broken gantry does. Worth checking whether your own QA would catch a workstation-side measurement error at all.

Playbook snippet

One step for your AI committee

Anchor: EU AI Act Article 10 (data governance).

Step: Add one clause to your next imaging AI RFP: the supplier discloses the composition of training and test data, including treatment phase, scanner mix and demographics, with performance reported by subgroup. Record refusal or silence as a risk item under Article 9 rather than treating it as acceptable.

Why: Training-set composition, not architecture, is what decides whether a model works on your use case, and it is the one thing a demo will never show you.

Worked example: Zielke J et al. trained a pediatric diffuse midline glioma segmentation model with post-treatment scans included and tested it externally on a prospective trial: whole-tumor Dice 0.90 vs 0.81 for the BraTS-PEDs 2024 challenge winner trained on pre-treatment data, with the largest gap on post-treatment scans, exactly where response assessment happens. AJNR Am J Neuroradiol, PMID 42648875, DOI 10.3174/ajnr.A9609.

From the field

RSNA published a news feature on a Radiology: Artificial Intelligence study testing 10 LLMs on 400 radiology board-style questions: prompts written to sound clinically plausible (authority, complexity, anchoring) cut accuracy by up to 21.1% on text and 44.9% on multimodal questions, with authority bias the most damaging. These are not adversarial tricks but recognizably human framings, exactly what a model embedded in clinical conversation will meet.

ACR reported that the Friends of NIBIB Coalition met with NIH officials to support imaging AI programs, notably PRIMED-AI, which will fund AI clinical decision support combining imaging with multimodal health data. ACR is also collecting member input for its response to the FDA's generative AI discussion paper, open for comments to 19 October.

On LinkedIn, Amine Korchi marked two milestones for ThinkSono, the AI-guided DVT ultrasound startup in his portfolio, arguing that AI's biggest impact will be redesigning how care is delivered rather than doing the same things better; in a second post he pushed back on the Swiss healthcare cost debate, noting a tariff hour is not an hour worked and that radiologist productivity reflects technology and organization. Curt Langlotz shared Modern Healthcare's feature on what radiologists want from the next generation of AI.

Daniel Pinto dos Santos amplified the ACR Data Science Institute's four questions for ACR-EuSoMII 2026 in Crete: is our AI performing as expected, who is responsible for monitoring it, how do we choose the next model, and is your governance governing. A ready-made self-test for any department's AI committee.

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.