Weekly AI digest
Radiology & medical imaging AI · week of 4 to 10 October 2026
199 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
Name the owner. Article 4 of the EU AI Act makes AI literacy an organizational obligation, and this week supplies both the map and the motive: EFOMP's new framework turns AI competence into an auditable professional ladder with six named gaps, and a carotid ultrasound reader study shows why it matters, with AI assistance lifting junior readers from near-chance baselines to usable but still below unassisted seniors.
This week, assign one named, competent owner to each deployed model and have them produce a single signed page: what the tool is cleared for, what your acceptance test showed, which patient subgroups and reader groups were checked, and what triggers suspension.
Regulation & AI Act watch
AIRS Medical's SwiftSight Body Composition cleared
FDA 510(k) for quantifying muscle, visceral and subcutaneous fat and liver fat from MRI, cleared on substantial equivalence with no product-specific published accuracy study, and arriving first through a self-pay longevity screening network. The question to ask: product-specific accuracy and repeatability against a dedicated PDFF protocol on your own scanners.MRI2CT's NextMR cleared for 3D spine bone visualization
510(k) for generating 3D bone visualization of the lumbar spine and sacrum from standard 1.5 T and 3 T diagnostic MRI in under 60 seconds; the vendor's own footnote says it is not intended to replace CT in general, and the cited validation is a 145-adult retrospective series without a named publication. The question to ask: performance with spinal hardware, deformity and transitional anatomy, and whether surgical planning still requires confirmatory CT.Avatar Medical Vision CE-marked with Barco's glasses-free 3D display
CE marking, announced at JFR 2026, for CT and MR visualization in 3D and VR for surgical planning; a visualization claim, not a detection claim, with no reader or outcome evidence in the release. The question to ask: the exact certified intended-use sentence and risk class, and the display QA the certification assumes.Bayer sells Blackford Analysis to Sopris Capital
The AI orchestration platform (130+ applications, 45+ vendors) changes owner for the second time in under four years and will be combined with oncology-workflow company Azra AI. The question to ask: which SLAs, data processing agreements and per-tool validation records survive the change of control.CARPL.ai to distribute HOPPR foundation models for local fine-tuning
A partnership that moves foundation-model fine-tuning into hospital procurement, with locally tuned report generation as the headline use case. The question to ask: who is the legal manufacturer, and who carries the regulatory obligations, for a model your team fine-tunes.
Evidence you can use
Selected from 199 papers indexed in PubMed this week.
Multimodal carotid ultrasound-based artificial intelligence model for predicting acute ischemic stroke risk in patients with type 2 diabetes mellitus
Academic Radiology, 9 October 2026Peritumoral ultrasound-driven multimodal AI model for predicting lymphovascular space invasion and prognosis in cervical cancer
Abdominal Radiology, 8 October 2026Radiology research platform: automating multicenter radiological research workflows using a cloud-native platform
European Radiology, 7 October 2026An open vision-language model for diverse medical applications (MedGemma)
Nature Medicine, 6 October 2026The medical physicist's role in the artificial intelligence era: an updated EFOMP curricular and professional framework
Physica Medica, 5 October 2026An inline-integrable image restoration model with ROI fine-tuning and slice-to-volume reconstruction for accelerated whole-brain vessel wall imaging
Magnetic Resonance in Medicine, 8 October 2026Decoupling multi-contrast super-resolution: self-supervised implicit re-representation for unpaired cross-modal synthesis
IEEE Transactions on Medical Imaging, 9 October 2026UPER: a unified polyhedral representation for universal multi-dimensional diffusion MRI image modeling
Medical Image Analysis, 6 October 2026
Post-market signal
Two compliance analyses published this week work through what the Digital Omnibus (Regulation (EU) 2026/1744, consolidated text since 27 July 2026) did to post-market monitoring under the EU AI Act. The implementing act that was to fix a mandatory monitoring-plan template by 2 February 2026 was never adopted and has been replaced: the Commission now owes guidance, including a template the amending act's recitals call voluntary, by 2 September 2027. Article 72's substantive duty is unchanged and formally applies from 2 August 2026, while the high-risk classification rules it attaches to now apply from 2 December 2027 (Annex III) and 2 August 2028 (Annex I, which covers MDR devices); how those dates fit together is openly flagged as unresolved.
The practical consequence for imaging: anyone waiting for an official template before starting monitoring is waiting for a voluntary document arriving late next year. Write the plan against Article 72(1) and (2) now, with named metrics, baselines, data sources, thresholds and a review cadence, and integrate it with the MDR post-market surveillance file your vendors already owe you. A device-focused reading of the amended text adds two concrete points: MDCG 2025-6 makes monitoring of interactions between AI systems the main genuinely new expectation, and Article 12 logging is a design requirement that must exist in the product, not only in a procedure.
The same signal came from the field this week: EuSoMII's annual meeting ran a dedicated post-market surveillance how-to session, and European informatics voices are explicitly calling for an EU equivalent of the ACR's Assess-AI registry.
Playbook snippet
One step for your AI committee
Anchor: EU AI Act Article 4 (AI literacy), applying since February 2025. The AI Act application calendar is under revision following the Digital Omnibus, so verify enforcement dates against the current text.
Step: Pick one deployed model. Produce a single signed and dated page: what it is cleared for, what your acceptance test showed, which patient subgroups and reader groups were checked, and what triggers suspension.
Why: A certificate proves attendance; a completed audit page proves the person can tell when the model is failing in their own population, which is what an inspector, and your own quality committee, will actually ask.
Worked example: Sun Y et al. (Academic Radiology, PMID 42855384, DOI 10.1016/j.acra.2026.09.036) show what the reader-group line of that page should contain: external-cohort AUC 0.939, senior readers improved from 0.868 to 0.929 with assistance, juniors from 0.548 to 0.659. If your acceptance test cannot show this split for your own readers, the audit page is not complete.
From the field
EuSoMII held its annual meeting in Heraklion on 9 and 10 October with a program that read like a governance syllabus: a dedicated post-market surveillance how-to session, the ACR Data Science Institute presenting its governance and monitoring programs including the Assess-AI registry, joint sessions with ESNR, SIIM, EFRS and MICCAI, and a keynote on European Health Data Space readiness.
JFR 2026 ran in Paris from 8 to 10 October with AI across the program. The emergency-imaging session carried the week's refrain: published performance never guarantees local performance, and drift monitoring belongs to deployment. One number stood out, from a Lancet-published randomized evaluation quoted there: 40% of practitioners had stopped using an AI tool within twelve months, and usage rate, not accuracy, determined its impact.
RSNA introduced an AI certificate program for imaging professionals on 4 October. Structured certification maps onto the Article 4 literacy obligation, but the distinction stands: a certificate documents training, while a completed model audit documents competence.
On LinkedIn, Daniel Pinto dos Santos amplified the ACR's Assess-AI message from EuSoMII, calling post-market surveillance urgently needed in the EU as well; Amine Korchi announced an investment in Vocca, AI voice agents for clinic phone lines, as a signal of where radiology-adjacent money is moving; and Matt Lungren explained his January departure from Microsoft with the launch of Oku Digital Health from Topcon, a bet that the next clinical AI platform grows out of eye care.
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.