Weekly AI digest

Three things that mattered

• Deep learning beats experts on breast, spine, and cardiac imaging

  Five peer-reviewed studies converge on the same finding: AI now outperforms expert clinicians. An MRI model classifies estrogen-receptor status non-invasively (AUC up to 0.923), a deep-learning model detects cervical cord compression on plain radiographs at 94.67% accuracy versus 69 to 71% for spine surgeons, and CMR-CLIP reaches 98.6% accuracy on hypertrophic cardiomyopathy.

• Three regulatory clearances and a new mammography model

  FDA 510(k) for the DeepHealth/RadNet prostate AI, CE marking for Philips SmartIQ coronary imaging and for Siemens AI Contouring of more than 200 structures, plus HOPPR's launch of a vision-language model for 2D mammography reporting. Validation and commercialization are moving in step.

• Governance and automated radiotherapy reach milestones

  The ACR and SIIM jointly approved the first formal practice parameter for AI in imaging and launched the Assess-AI quality registry. At ESTRO 2026, the ARCHERY trial showed AI can design reference-quality chemoradiotherapy plans for cervical cancer in more than 95% of cases.

Peer-reviewed

• MRI-based deep learning for triple-class classification of estrogen-receptor expression in breast cancer

  European Radiology · multicenter consortium, 6 institutions, 20 May 2026

• Deep-learning model versus spine surgeons for detecting cervical cord compression on radiographs

  European Spine Journal · spine surgery institution, 22 May 2026

• Explainable AI in oncologic imaging: a scoping review of methods, modalities, and clinical integration

  JMIR Medical Informatics · international research consortium, 20 May 2026

• RetCond: a conditional diffusion model for self-explainable multiclass classification of fundus images

  Diagnostics (MDPI) · Canadian university, 22 May 2026

• CMR-CLIP: contrastive image-language pretraining for a cardiac MRI embedding model with zero-shot capability

  Nature Communications · Cleveland Clinic / Carnegie Mellon University, 21 May 2026

Industry & regulation

• DeepHealth/RadNet earns FDA 510(k) and CE marking for prostate MRI lesion-detection AI

  20 May 2026. Multiparametric MRI analysis showing 27% improved detection of clinically significant lesions, 65% lower inter-reader variability, and 37% shorter workflow time. First DeepHealth prostate AI tool to obtain both clearances.

• Philips SmartIQ receives CE marking for ultra-low-dose coronary imaging

  18 May 2026, presented at EuroPCR 2026 (Paris). AI imaging on the Azurion platform that cuts radiation by more than 50% versus ClarityIQ. The randomized RADIQAL trial is ongoing at 60% enrollment.

• Siemens AI Contouring for Eclipse receives CE marking

  Announced at ESTRO 2026 (Stockholm). Automatic contouring of more than 200 predefined anatomical structures on CT and MRI, integrated into the Varian Eclipse planning system for radiotherapy.

• Bracco and ACIST Medical Systems clear ACIST Pro via FDA 510(k)

  18 May 2026. Contrast-management system for interventional cardiovascular procedures, aimed at optimizing contrast injection, reducing volumes, and improving reproducibility.

• HOPPR launches a vision-language model for structured 2D mammography reporting

  19 May 2026. Trained on more than 200,000 studies from multiple sites to generate structured reports from raw mammographic images. Joins HOPPR's VLM portfolio for 3D mammography and chest radiographs.

Conferences & voices

ESTRO 2026 (Stockholm, 15 to 19 May) put AI and automation at the center of radiation oncology, with the ARCHERY trial results and the Siemens AI Contouring CE marking. At ACR 2026 (Washington DC), the ACR and SIIM jointly approved the first formal practice parameter for clinical AI in imaging and launched the Assess-AI post-deployment quality registry, while EuSoMII confirmed its annual meeting for Heraklion, Crete, on 9 to 10 October 2026. Among KOL voices, Woojin Kim (HOPPR, ACR Data Science Institute) argued that generative AI for reports will require more radiologists, not fewer, framing the risk as inadequate deployment rather than replacement. Curtis Langlotz (Stanford AIMI) outlined a plan for a radiology foundation model trained on Stanford's full archive of roughly 2 petabytes.

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 practicing radiologist. Free, no vendor sponsorship.

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