Weekly AI digest
Radiology & medical imaging AI · week of 16 to 22 August 2026
210 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
If chest radiography AI is live or planned in your department, schedule a local multi-reader check this month and score accuracy per finding type, not pooled. In a prospective crossover study of four commercial tools, readers became faster and more confident while accuracy stayed flat and fell for effusions and nodules through added false positives. Speed and confidence are exactly what a demo shows you; neither is evidence of accuracy.
The same week's osteonecrosis meta-analysis measured the discount to apply to internal validation claims: external testing roughly halved the diagnostic odds ratio. Test the claim you will actually rely on, in your own population, before it shapes reader behavior.
Regulation & AI Act watch
FDA opens comment period on generative AI-enabled devices
CDRH published a discussion paper (docket FDA-2026-N-7874, comments due 19 October 2026) outlining a two-axis risk framework, competency-based premarket evaluation and risk-proportionate postmarket monitoring. The consequential idea for deployers: the agency is weighing whether to accept greater premarket uncertainty in exchange for heavier reliance on postmarket monitoring, which would shift validation burden toward the institutions that deploy these tools. The ACR is preparing a coordinated radiology response.Radiology Partners petitions FDA on vision-language models
Mosaic Clinical Technologies filed a citizen petition asking the FDA to clarify whether commercially distributed imaging vision-language models, including those marketed for later institutional fine-tuning, are medical devices requiring premarket review. A major AI developer asking for more scrutiny of its own product category is notable; so is the petition's plain statement that deploying practices currently bear the validation burden for un-reviewed foundation models.Cortechs.ai NeuroQuant PET cleared
510(k) clearance (K261916) for automated quantification and structured reporting of PET in cognitive impairment, generating SUVR and Centiloid values for three amyloid tracers. Evidence level is vendor materials; the one question to ask first: what is the test-retest variability of Centiloid output on your own scanners before results feed dementia pathways.CureMetrix cmAngio 2.0 cleared with BAC quantification
Adds quantification, age-based percentile and prevalence context for breast arterial calcification on screening mammograms, with vendor-reported AUC 0.98 to 0.99. The one question to ask first: who acts on a high BAC percentile and through which funded cardiology pathway, since notification laws are starting to make BAC a reportable result.Distribution deals, not new evidence
Azra AI and 4DMedical signed an OEM and reseller partnership joining report-mining patient identification with quantitative lung imaging analytics, and Nanox AI signed an exclusive UK reseller agreement with Vertec Scientific for its HealthOST bone solution. Both are commercial reach, not new clearances; the local validation burden is unchanged. A PLOS Digital Health analysis published the same week frames the wider gap: of 1,357 FDA-cleared AI devices, only three have been evaluated against patient outcomes.
Evidence you can use
Selected from 210 papers indexed in PubMed this week.
Large-scale AI-guided liver malignancy diagnosis: multicenter study and a single-arm trial
Nature Medicine, 19 August 2026Artificial Intelligence-Assisted Chest Radiography: A Prospective Crossover Multi-Reader Study on Diagnostic Performance and Workflow Efficiency
Academic Radiology, 21 August 2026Diagnostic Accuracy of Medical Imaging-Based Artificial Intelligence for Osteonecrosis of the Femoral Head: Systematic Review and Meta-Analysis
Journal of Medical Internet Research, 20 August 2026Multicenter External Validation of an AI-Based Funduscopic Carotid Atherosclerosis Score and Assessment of Its Association With Coronary Artery Calcification
JMIR Medical Informatics, 19 August 2026Automated deep learning-based segmentation and volumetric analysis of meningiomas
Neuroradiology, 20 August 2026A stacking model for AI-assisted diagnosis of suspected pituitary microadenomas on non-contrast T1COR MRI: a multicenter reader study on bridging the experience gap
Neuroradiology, 20 August 2026A comprehensive framework for multi-class prostate cancer classification using biparametric MRI: a multi-center retrospective study
Insights into Imaging, 17 August 2026Motion Artifact-Aware Self-Supervised Representation Learning for 3D Brain MRI Motion Artifact Reduction
IEEE Transactions on Medical Imaging, 17 August 2026
Post-market signal
An FDA Class 2 recall covering nearly 11,000 Philips Azurion R2.2.10 interventional fluoroscopy systems was reported this week. Two software issues can degrade image quality or disable geometry movements, with potential for delayed therapy or procedural complications in exactly the patients least able to tolerate delay. Philips sent Urgent Medical Device Correction letters in mid July with interim workarounds, a cold-restart procedure, pending a software update.
The lesson for imaging departments is not about this vendor. Software regression in a mature, widely deployed platform is a live failure mode, and the incident path from a technologist noticing artifacts to the vendor complaint file is the mechanism that makes a recall like this visible. Verify that path exists in your department and that people use it.
Playbook snippet
One step for your AI committee
Anchor: EU AI Act Article 14 (human oversight).
Step: Before deployment, record each reader's baseline on a fixed local case set: reading time and detection or marking rate per finding type. Repeat at 3 months with the tool on, and report per-reader deltas; pooled averages hide the effect.
Why: Human oversight is only meaningful if you know what the tool does to each specific human. A tool that speeds up one reader and degrades another looks neutral in the average, and the readers themselves cannot feel the difference: confidence rises whether or not accuracy does.
Worked example: Lemke T et al. tested four commercial chest radiography AIs with five readers over 1200 consecutive patients: accuracy fell for effusions and nodules in specific reader-tool pairings while three of five residents got faster and four of five felt more confident. Only the per-pairing analysis surfaced this. Academic Radiology, PMID 42629289.
From the field
The RSNA R&E Foundation approved more than 4.5 million dollars across 79 grants for 2026, spanning AI-assisted detection of perihilar cholangiocarcinoma, a 5-minute breast MRI protocol, whole-body MRI to study GLP-1 effects, and the foundation's first funded project in Argentina, an open-source AI-assisted training simulator. Applications for most 2027 grants open in October.
One day after the FDA published its generative AI discussion paper, the ACR announced it is preparing a coordinated radiology response through the Data Science Institute and invited members to contribute before the 19 October deadline.
Woojin Kim highlighted new behavioral evidence that AI advice suppresses the willingness to say "I don't know": across five experiments with 3,132 participants, declining to answer fell from 44% to 3% once AI advice was available, even though the advice was engineered to be wrong. His framing for radiology: AI suggestions may alter the metacognitive threshold at which clinicians decide they know enough to answer.
Curt Langlotz amplified the FDA paper release and relayed the call for papers for AIM-NeurIPS 2026, the first NeurIPS workshop on agentic intelligence for medical imaging. Amine Korchi announced a new European Journal of Radiology Artificial Intelligence paper with Jan Beger and Christoph Agten, "Three futures for the diagnostic radiologist", a structured disagreement laying out expansion, concentration and stratification scenarios for the specialty.
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.