← All issues

Weekly AI digest

Radiology & medical imaging AI · week of 30 August to 5 September 2026

239 peer-reviewed papers · 6 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 the monitoring layer itself came under the microscope. The ACR's Assess-AI registry, the first national imaging AI registry, monitors AI outputs by having an LLM extract findings from signed reports, and its authors admit the extractor's own accuracy is not yet independently validated. Read that next to CIDER, where an extraction LLM was held to per-field accuracy and repeated-run reproducibility standards, and next to a TI-RADS study where a prompt change alone moved scoring accuracy by 28 points.

The action for a radiology leader: inventory every place an LLM silently turns text into data in your department, including inside your monitoring tools, and give each one the same acceptance test you would give a diagnostic model, with the prompt version-locked and test-retest agreement measured before its output counts as evidence.

Regulation & AI Act watch

Evidence you can use

Selected from 239 papers indexed in PubMed this week.

Post-market signal

The stronger signal this week is that post-market monitoring is becoming a certified product component rather than a paper obligation. Vara's autonomous screening CE mark rested on ATMON, a real-time supervision system grounded in seven years of continuous real-world monitoring, and the company will license ATMON independently so providers could wrap it around other autonomous tools. The regulator accepted autonomy only with a live safety net attached; if a vendor proposes any autonomous operation, ask what their equivalent of ATMON is and who receives its alarms.

The second signal is a model of post-deployment measurement from Erasmus MC, published in Radiology on 1 September: 19,190 chest CT exams the year before deploying a commercial nodule tool versus 20,133 the year after. Median reporting time fell 14.6 percent, with the largest gains for ECG-gated exams and thoracic subspecialists, while emergency department exams got 7.1 percent slower. Aggregate benefit coexisted with a measurable subgroup harm signal, visible only because the department segmented its own routine data.

Playbook snippet

One step for your AI committee

Anchor: EU AI Act Article 15 (accuracy, robustness and cybersecurity).

Step: Add reproducibility as a separate acceptance test for any LLM-based tool: run the same 30 of your own documents or cases through it twice, days apart, at the vendor's production settings. Report test-retest agreement alongside accuracy and set a threshold below which the tool does not go live, with the prompt and model version recorded in the acceptance file.

Why: A model that answers differently on Tuesday is not deployable however accurate it was on Monday, and accuracy testing alone will never surface this.

Worked example: Posta M et al. validated the CIDER extraction pipeline on 2073 Hungarian histopathology reports with exactly this design: robustness across temperatures 0 to 2.0 and technical reproducibility across 3 independent runs at temperature 0.1, alongside per-field exact-match accuracy from 99.5 percent (sex) down to 78.1 percent (tumor size). J Med Internet Res, PMID 42686197, DOI 10.2196/95780.

From the field

The ACR spent the week in the payment arena: comments filed on the CY2027 HOPPS proposed rule, cautioning against site-neutral imaging payment and pressing CMS on how Software as a Medical Service gets categorized and paid, plus three new draft QCDR measures announced for 2027 MIPS. How algorithmic services are reimbursed in the outpatient setting is being negotiated now, ahead of the final rule this fall.

RSNA's inaugural Across Africa: Advancing Cancer Imaging course is running 4 to 5 September in Cape Town, combining cancer imaging clinical sessions with capacity-building case studies from Egypt, Tanzania, Uganda and Kenya, and closing on AI innovation in global cancer care.

On LinkedIn, Curt Langlotz greeted the first autonomous breast screening approval with two words, "It's happening...", a compact marker that the debate has shifted from whether to under what supervision. Woojin Kim previewed his KCR 2026 lectures on foundation models and agentic AI and amplified the new ACR buyer's guide on calculating the AI errors that matter to your own population.

Pranav Rajpurkar reported the ReXGroundingCT Challenge at MICCAI 2026 heating up: 36 teams, top Dice up from 0.332 to 0.367, ten days left. Grounding, making a model show which pixels support each generated sentence, is the capability report-drafting AI needs before its output can be audited rather than merely read. Amine Korchi, meanwhile, announced a move into long-form video for the ideas that need more room than a feed post.

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.