← All issues

Weekly AI digest

Radiology & medical imaging AI · week of 20 to 26 September 2026

194 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

This week draws the line between accuracy evidence and deployment evidence unusually clearly. A Nature Medicine screening model arrives with 12-center validation on 80,612 patients, real-world calibration and a prospective cohort, while the week's meta-analyses and benchmarks, LVO triage on noncontrast CT, cerebral microbleed detection, LLM decision support, all stop at retrospective accuracy, and say so.

The action for a radiology leader: add one line to every AI proposal your committee reviews, classifying its evidence as accuracy-based or workflow-based, and refuse to approve an unsupervised or workflow-changing role on accuracy-only evidence. A bounded human-in-the-loop role with a prospective monitoring plan is the correct ceiling until the workflow evidence exists.

Regulation & AI Act watch

Evidence you can use

Selected from 194 papers indexed in PubMed this week.

Post-market signal

The strongest post-market evidence of the week came from endoscopy. Cosmo announced results of the CADeNCE study (Dominitz et al., Gastroenterology) on 21 September: a cluster-randomized quality-improvement evaluation of the GI Genius colonoscopy CADe system across more than 334,000 procedures by 816 endoscopists at 139 US Veterans Health Administration facilities. CADe availability was associated with 22 percent higher odds of adenoma detection (adjusted OR 1.22; adenoma detection rate 50.7 to 54.9 percent), with benefit across all baseline detection levels. This is what post-market evidence should look like: randomized, national scale, run on the deployed system in routine practice. One detail belongs in every monitoring plan: the study ran on an earlier software generation than the currently marketed version, a reminder that post-market evidence attaches to a version, not a product name.

A second item is commentary rather than an imaging post-market event: a critical-care practice newsletter (25 September) pulled together the emerging literature on decommissioning deployed clinical AI, including radiology's JACR off-boarding framework and an NEJM AI analysis of the measurement trap in which a model that changes outcomes corrupts its own monitoring statistics. The transferable points: require a written exit and transition plan before any tool goes live, track vendor end-of-support dates by name, and treat silent version updates as the most common and least visible exit mode.

Playbook snippet

One step for your AI committee

Anchor: EU AI Act Article 12 (record-keeping) and Article 19 (automatically generated logs).

Step: At procurement, ask the vendor in writing what the system logs, for how long, in what format, and whether you can export it without their involvement. If the answer is a dashboard, the answer is no. Add the log-export clause to the contract before signature, not after go-live.

Why: Every post-market obligation downstream assumes you can reconstruct what the model saw and said on a given date. Retrofitting logging after deployment is a contract renegotiation, not a configuration change.

Worked example: Zhou J et al. show what retained case-level outputs buy you: EAGLE's real-world calibration on 35,402 scans cut false positives 72.7 percent while preserving sensitivity, an adjustment only possible because outputs across three centers could be re-analyzed after the fact. Demand the same reconstruction capability from any tool you deploy. Nature Medicine, 22 September 2026.

From the field

The societies spent the week on governance infrastructure. The Royal College of Radiologists published its reading of the UK National AI Commission's recommendations on healthcare AI regulation, confirming it is working on proposals for a national post-deployment monitoring system. RSNA announced a RadLex initiative to standardize imaging series naming, the unglamorous metadata layer whose inconsistency silently breaks deployed AI pipelines.

Reimbursement moved too. At the September AMA CPT Editorial Panel, ACR and five partner societies requested a Category III code for a five-year AI-derived breast malignancy risk score from screening mammography, the first step toward reimbursed AI risk stratification. The ESR Patient Advisory Group placed cardiovascular imaging at the center of the EU Safe Hearts Plan.

Among the voices, Nina Kottler posted a widely shared three-part thesis: responsible AI requires lifecycle governance because validation at deployment is not enough, scaled governance requires payment models that fund monitoring infrastructure rather than only the AI product, and AI expands radiology's role only if the specialty owns validation and clinical meaning. Curt Langlotz amplified Johns Hopkins' new NIH R01 for detecting three abdominal cancers on routine CT, the same early-detection territory as this week's Nature Medicine esophageal screening paper.

Woojin Kim lectured at Brown on generative and agentic AI in medical imaging, and Amine Korchi reminded his readers that a radiologist's workload is a 100-line worklist invisible to everyone outside the reading room, the capacity strain that drives both AI adoption and burnout.

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.