← All issues

Weekly AI digest

Radiology & medical imaging AI · week of 11 to 17 October 2026

228 peer-reviewed papers · 4 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

Validate the use, not the tool. This week the same mammography AI swung from modest to large gains depending on how it was wired into the read; adult chest models failed on neonates until they were adapted with 50 local radiographs per site; and the largest prostate MRI AI study to date matched radiologists on average while leaving 28% of sites below an AUC of 0.80.

For every deployed model, write a single line naming the integration policy, who sees the output and when it can override, and the patient and reader population your acceptance test actually covered. Treat any use outside that line, a new scenario, a new population, a new site, as a new deployment that needs its own acceptance test and monitoring baseline.

Regulation & AI Act watch

Evidence you can use

Selected from 228 papers indexed in PubMed this week.

Post-market signal

Two publications this week move post-market surveillance from aspiration toward tooling. In npj Digital Medicine, AEGIS describes an operational infrastructure for governing adaptive medical AI under FDA Predetermined Change Control Plans and EU post-market surveillance: every candidate model update is routed into APPROVE, CONDITIONAL APPROVAL, CLINICAL REVIEW or REJECT, with an independent ALARM state for a released model at risk. On real cross-site data, batch-wise drift detection reached ROC-AUC 1.00 on batches of about 500 patients with no false alarms. It is a proof of concept covering covariate shift only, but the four-outcome taxonomy and the batch-wise drift gate are a usable skeleton for an Article 72 monitoring file today.

A WAIRR explainer published October 14 reads the drift evidence critically rather than reverently. The Grampian mammography analysis shows a manufacturer threshold set elsewhere would have flagged 48.3% of exams against 5.4% for the first human reader, and a scanner software update tripled the flag rate after site calibration. The Mayo Clinic CTPA disagreement monitor shows the asymmetry that matters: a 2 to 3% fall in specificity produced a two to threefold rise in discrepancies, while a 10% sensitivity drop was needed for the same signal. Disagreement monitors catch extra false alarms quickly and missed findings slowly, which is the more dangerous failure to miss.

Playbook snippet

One step for your AI committee

Anchor: EU AI Act Article 14 (human oversight).

Step: Before a new tool goes live, record each reader's baseline on a fixed local case set: reading time plus detection or marking rate. Repeat at 3 months with the tool on and report per-reader deltas, never only the pooled average.

Why: Pooled averages hide what a tool does to specific humans: assistance can lift junior readers, slow senior readers, or move two readers in opposite directions, and oversight is only real if you know which is happening in your reading room.

Worked example: Zhou Y et al. (Journal of Imaging Informatics in Medicine, PMID 42870161, DOI 10.1007/s10278-026-02352-w): with an interpretable MRI model for endometrial stromal sarcoma, the junior radiologist's specificity rose from 46.4% to 82.1%, exactly the per-reader delta an Article 14 oversight file should contain. This week's mammography scenario study (PMID 42867728) makes the same point at policy level: the integration scenario, not the model, produced the gain.

From the field

SIIM's Enterprise Imaging webinar series hosted Sectra on October 14 for a live demonstration of autonomous AI clearing normal chest exams from the worklist, the capability acquired with Oxipit, with the validation and safety thresholds applied before deployment on the agenda. Autonomous normal-clearing is the sharpest test of local validation discipline there is, since no human reads what the model clears.

RSNA published participant feedback on its Imaging AI Certificate Program on October 16: completers report stronger skills in evaluating, implementing and monitoring AI tools, with a Neuroradiology certificate due late 2026. The Journal of Imaging Informatics in Medicine also released its SIIM 2026 abstract supplement, a useful scan of what imaging informatics teams are building under governance, privacy and interoperability constraints.

On LinkedIn, Pranav Rajpurkar announced the Nature Biomedical Engineering study behind this week's flagship selection, 30 neonatal ICUs adapting an adult model with 50 radiographs each, and framed it plainly: newborns are among the patients medical AI serves least. Woojin Kim shared the ACR Data Science Institute's EuSoMII wrap-up, where he presented the ACR's work on imaging AI quality and performance monitoring, including the Assess-AI registry.

Amine Korchi argued that many radiology inefficiencies share one root cause, poor cross-functional understanding between radiologists, technologists and secretaries, and that operational excellence takes designed processes and training, not more individual effort. A reminder that AI lands on top of exactly these processes.

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.

By subscribing you agree to receive the weekly digest by email. Unsubscribe anytime. No sharing with third parties.