← All issues

Weekly AI digest

Radiology & medical imaging AI · week of 9 to 15 August 2026

187 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 mammography AI is live in your department, split recall monitoring by pathway this month. A year-long alternating-month deployment doubled cancer detection without slowing reading, but abnormal interpretations rose only in diagnostic mammography, 18.7% versus 12.1%, while screening stayed flat. A single pooled recall metric would have hidden that shift.

Add abnormal interpretation rate by pathway to your monthly monitoring file; the same file becomes the seed of your EU AI Act Article 72 post-market record.

Regulation & AI Act watch

Evidence you can use

Selected from 187 papers indexed in PubMed this week.

Post-market signal

SIIM published a recap of its webinar on post-production monitoring of deployed radiology AI, and it reads as a field report on what Article 72-style obligations look like in practice. Penn Medicine runs a governance committee spanning demo approval through post-deployment monitoring; Dasa tracks drift on an in-house MRI denoising model by watching input voxel distributions and input-output similarity over time; and Greensboro Radiology has declined FDA-cleared tools over poor real-world performance, monitoring deployed algorithms by comparing predictions against radiologist impressions extracted from reports. The shared message: clearance is the start of the evidence obligation, not the end.

A Clinical Imaging survey of 215 Society of Breast Imaging members adds the deployer-side view. About half worked in organizations with breast detection AI live, yet only a minority reported measurable improvement in callback rate, biopsy rate or burnout; recall rates fell around 35% among current users against non-user expectations, unnecessary biopsies only about 9%. Adoption is running ahead of measured benefit, which is exactly why departments should track their own deltas rather than projected ones.

Playbook snippet

One step for your AI committee

Anchor: EU AI Act Article 72 (post-market monitoring); ISO/IEC 42001 performance evaluation clause.

Step: Sample 20 AI outputs per month against the final signed report and track one number: the concordance rate. Plot it monthly and act on the trend, not on any single dip.

Why: Drift is invisible in aggregate accuracy and visible in a monthly line; a series started now is history you will already own when the monitoring obligation bites.

Worked example: Lakhani P et al. audited 400 AI-drafted DXA reports against the source images, not the prior reports, across four sites: 99.9 to 100% numerical accuracy, and the audit surfaced a 45% completeness rate in the human baseline at community sites. JACR, PMID 42595274, DOI 10.1016/j.jacr.2026.08.007.

From the field

The ACR published its detailed summary of the CMS FY2027 IPPS final rule on 13 August: a 2.3% hospital payment update, roughly $779 million in new-technology case payments, and the end of the alternative NTAP pathway from FY2028, after which every technology seeking add-on payment must demonstrate substantial clinical improvement. The evidence bar for US inpatient reimbursement just moved up.

The American Board of Radiology set out its AI position in a 12 August update: no AI in certification decisions or exam content, human experts retain scoring, and governance structures come before any adoption. A certification body publishing its human-oversight boundary is itself a governance signal for the specialty.

Curt Langlotz spent the week amplifying work on confidence-based control of clinical report generation, where an operating point on the false-positive curve gates what an LLM writes into the draft, alongside Stanford AIMI's pediatric echocardiography model in Circulation and the Nature Health analysis of 617,827 health conversations with Microsoft Copilot.

Daniel Pinto dos Santos flagged the joint EuSoMII and MICCAI session on the future of AI in medical imaging at October's EuSoMII Annual Meeting in Heraklion, themed on safe integration of AI tools in practice, while Amine Korchi pointed to voice AI agents for practice administration as the segment of medical AI with the lowest deployment friction.

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.