Weekly AI digest

Three things that mattered

• An autonomous AI agent surpasses physicians in diagnostic accuracy

  A Nature paper published on 17 June presents MIRA, an AI agent integrated with electronic health records that outperforms physicians in diagnostic accuracy including imaging interpretation and treatment planning. The work represents a conceptual shift from decision support toward autonomous clinical reasoning in medicine.

• FDA formalizes adaptive AI with a PCCP final order for imaging software

  On 17 June the FDA published a final order classifying quantitative ML imaging software with Predetermined Change Control Plans as Class II devices under 21 CFR Part 892. Manufacturers can now update their algorithms post-market under pre-approved protocols without submitting a new 510(k) for each update.

• Two studies confirm zero-shot LLMs fall short in core radiology tasks

  Published in European Radiology and Scientific Reports on 17 and 19 June, two independent evaluations found that GPT-4.1, GPT-4o, Claude 3.5 Sonnet, and Gemini 1.5 Pro all failed to meet clinical thresholds in zero-shot settings: the best LLM detected physiologically impossible CT report errors at 16.3%, and all three multimodal LLMs performed near chance on pediatric wrist fracture detection against a domain-trained CNN with AUROC 0.905.

Peer-reviewed

• AI-assisted volumetric response criteria in pleural mesothelioma (ARTIMES)

  The Lancet Oncology, Netherlands Cancer Institute and 14 international cohorts, 17 June 2026

• Toward autonomous medical artificial intelligence agents (MIRA)

  Nature, multicenter, 17 June 2026

• GPT-4.1 and Llama 3.3 70B fail to detect clinically relevant errors in radiology reports in zero-shot evaluation

  European Radiology, TU Munich and University of Basel, 19 June 2026

• Zero-shot multimodal LLMs underperform a domain-trained CNN baseline in pediatric wrist fracture detection

  Scientific Reports, Carl von Ossietzky University Oldenburg, 17 June 2026

• Revisiting perfusion in the right ventricle: AI-enhanced quantification on PET/CT

  Journal of Nuclear Medicine, Cedars-Sinai and 10 partner sites, 18 June 2026

• Segmenting with confidence through uncertainty quantification for brain tumor imaging

  NPJ Digital Medicine, UCSF Center for Intelligent Imaging, 19 June 2026

• The AI implementation gap in trauma radiography: standalone versus discretionary AI-integrated fracture detection

  European Radiology Experimental, Semmelweis University Budapest, 16 June 2026

• Predicting early-stage breast cancer disease-free survival from multimodal deep learning on MRI and clinical data

  NPJ Breast Cancer, Guangdong Provincial Hospital and Southern Medical University, 17 June 2026

Industry & regulation

• FDA final order establishes PCCP pathway for adaptive ML imaging software

  Quantitative imaging software with a Predetermined Change Control Plan is now Class II under 21 CFR Part 892, allowing manufacturers to update algorithms post-market without a new 510(k) submission for each change.

• Midjourney Medical announces full-body ultrasound scanner in partnership with Butterfly Network

  The deal involves up to 74 million dollars over five years, but the device has no FDA clearance, the claimed 60-second scan time is not met by the current prototype, and no clinical validation data have been published.

• TA Associates invests in AIRS Medical

  SwiftMR, the company's deep learning MRI reconstruction product, operates in more than 1,700 institutions across 40-plus countries and processes approximately 6 million exams per year.

• FDA AI device tracker reaches 1,524 authorizations with 76% in radiology

  The first quarter of 2026 added 92 new authorizations, a 28% increase over the fourth quarter of 2025, with radiology continuing to lead all medical specialties.

• Rad AI and Yale New Haven Health deploy ambient radiology reporting at scale

  The integration covers 16 outpatient sites and 5 hospital campuses, with a 12% efficiency gain recorded in the first week of deployment.

• Radiology AI devices carry 52% higher recall risk, JAMA Network Open reports

  An analysis of 903 FDA-authorized AI devices found that 5% were recalled, with radiology devices carrying the highest risk, most often linked to insufficient clinical study data in the submission dossier.

Conferences & voices

SIIM 2026 in Pittsburgh marked a turning point from AI experimentation to routine operational deployment, with "workflow intelligence" emerging as the defining framing for the next phase of clinical AI integration. The joint ACR-SIIM Practice Parameter on imaging AI was introduced as the first sector-wide governance standard covering deployment, validation, and ongoing monitoring of AI systems. Woojin Kim was named Fellow of SIIM in recognition of his contribution to imaging informatics. Eric Topol noted on LinkedIn on 12 June that general-purpose LLMs substantially outperformed specialized medical models in a rigorous head-to-head comparison, a finding relevant to AI tool selection strategies in radiology. Yashbir Singh of Mayo Clinic stated that human oversight in AI-augmented radiology is a structural requirement rather than a transitional measure, a position echoed by the radiology community's critical response to Midjourney Medical's unvalidated full-body ultrasound claims.

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.