Weekly AI digest

Three things that mattered

• Agentic AI crosses a clinical threshold

  MIRA, published in Nature on 17 June, shows an EHR-integrated AI agent matching or exceeding physicians on diagnostic accuracy, including imaging. The autonomy question moves from "if" to "under what governance."

• Claims vs reality in consumer imaging

  Midjourney Medical and Butterfly Network announced a whole-body ultrasound scanner. Documented gap between a 60-second claim and roughly 20 minutes in the prototype. A reminder to benchmark vendor claims against real-world performance.

• LLM limits in radiology confirmed

  Two independent studies (European Radiology, Scientific Reports, 17 to 19 June) show zero-shot general LLMs underperform on radiology error detection and fracture tasks versus trained models.

Peer-reviewed

• ARTIMES: AI-assisted volumetric response criteria in pleural mesothelioma

  The Lancet Oncology · Netherlands Cancer Institute, 14 international cohorts

• MIRA: toward autonomous medical AI agents integrated with the electronic health record

  Nature, 17 June 2026

• GPT-4.1 and Llama 3.3 70B fail to detect clinical errors in radiology reports (zero-shot)

  European Radiology · TU Munich / University of Basel, 19 June 2026

• Multimodal zero-shot LLMs vs a trained CNN for pediatric wrist fracture detection

  Scientific Reports · University of Oldenburg, 17 June 2026

• AI-assisted right-ventricular quantification in myocardial perfusion PET/CT

  Journal of Nuclear Medicine · Cedars-Sinai, multicenter

• Meningioma segmentation with calibrated uncertainty on brain MRI

  npj Digital Medicine · UCSF Center for Intelligent Imaging, 19 June 2026

• The AI implementation gap in trauma: autonomous vs discretionary use for fractures

  European Radiology Experimental · Semmelweis University, 16 June 2026

• Multimodal deep learning for disease-free survival in early breast cancer (MRI plus clinical)

  Guangdong Provincial Hospital / Southern Medical University

Industry & regulation

• FDA final order: PCCP classification of ML imaging software as Class II

  17 June 2026. Predetermined change control plans move further into mainstream regulatory practice.

• FDA AI device base: 1,524 authorizations, 76% in radiology

  Q1 2026 up about 28%. Radiology remains the dominant cleared-AI specialty.

• AI device recalls: radiology shows 52% higher risk

  JAMA Network Open. Post-market surveillance and monitoring matter.

• Rad AI and Yale New Haven Health: large-scale ambient reporting

  Announced at SIIM. Deployment, not pilot.

• TA Associates acquires a stake in AIRS Medical (SwiftMR)

  Continued private capital interest in radiology AI.

Conferences & voices

SIIM 2026 (Pittsburgh) framed AI's move from experimentation to operations, alongside the first joint ACR-SIIM Practice Parameter for AI in imaging. Eric Topol highlighted rigorous LLM benchmarking; Mayo Clinic's Yashbir Singh reiterated that human oversight is a structural necessity, not a temporary measure.

Source articles are indexed in PubMed with verified DOIs. Manuscript-stage work is excluded. The full weekly report is produced in French.