Weekly AI digest

Three things that mattered

• Large language models take over report structuring

  Several papers converged on LLMs for report-level tasks: a hybrid rules plus LLM workflow (Falcon3-10b) reached 99.8% accuracy and an F1 of 0.99 on 2,923 pulmonary embolism CT reports, beating radiologists, while an offline LLaMA-3.3 70B classified oncology CT reports by RECIST criteria with no fine-tuning or external data transfer. Woojin Kim named automated report generation the dominant 2026 trend.

• Three FDA 510(k) clearances in one week

  Subtle Medical cleared SubtleHD(PET) (K254013) for AI PET image enhancement with up to 75% faster acquisition, Cercare Medical cleared a CBCT Perfusion solution (K253831) for real-time perfusion in the neuro-interventional suite, and DeepHealth secured FDA and CE milestones for Prostate AI, LumbarMR, and its Neuro Suite.

• Explainability and reasoning move to the foreground

  Explainable AI cut automation and anchoring bias in mammography roughly in half (from about 35% to about 17%, p<0.001) and lifted accuracy from 86.2% to 90.1%. In parallel, Stanford AIMI released CheXthought, a dataset of 501 radiologists from 71 countries capturing expert visual reasoning to train AI to reason like a radiologist.

Peer-reviewed

• Cognitive bias in AI-assisted mammography interpretation: a simulation study with explainable AI

  European Radiology · European Institute of Oncology IRCCS, Milan, 29 May 2026

• ChestX-Reasoner: advancing radiology foundation models with step-by-step verification from routine reports

  Communications Medicine · Shanghai Jiao Tong University / Shanghai AI Laboratory, 27 May 2026

• Who labels best? Radiologists, rules, or large language models for pulmonary embolism CT reports

  European Radiology Experimental · Heidelberg University, 27 May 2026

• AI-MIRACLE: AI and multiparametric MRI to predict neoadjuvant immunotherapy outcomes in muscle-invasive bladder cancer (PURE-01)

  European Urology Oncology · IRCCS San Raffaele, Milan / Memorial Sloan Kettering Cancer Center, 29 May 2026

• Machine-learning model from pre-treatment cardiac MRI to predict cancer-therapy-related cardiac dysfunction

  JACC: Cardiovascular Imaging · Toronto General Hospital / University of Calgary, 28 May 2026

• Improved resolution and image quality in musculoskeletal MRI on a 1.5T scanner with deep-learning reconstruction

  European Radiology Experimental · Istituti Clinici Zucchi, Monza / ASST Papa Giovanni XXIII, Bergamo, 28 May 2026

• Automated RECIST tumor response classification with prompt-guided large language models (offline LLaMA-3.3)

  Scientific Reports · Technical University of Munich, 27 May 2026

Industry & regulation

• Subtle Medical clears SubtleHD(PET) via FDA 510(k) (K254013)

  27 May 2026. AI PET image enhancement for PET/CT and PET/MR, with up to 75% faster acquisition on existing scanners. Supports all FDA-approved radiotracers beyond 18F-FDG, with adjustable denoising and CT-informed reconstruction. More than 1,300 scanners already run the predecessor, SubtlePET.

• Cercare Medical clears CBCT Perfusion via FDA 510(k) (K253831)

  28 May 2026. First qualitative cerebral perfusion assessment directly from C-arm data during neuro-interventional procedures, without moving the patient. Delivers rCBF, rCBV, MTT, Tmax plus advanced biomarkers (OEF, CMRO2, CTH); clinical validation by 13 neuroradiologists at Inselspital, Bern.

• DeepHealth secures FDA and CE milestones for Prostate AI, LumbarMR, and Neuro Suite

  Announced 19 May 2026. CE marking for Brain Health and Brain Age (Neuro Suite) and for LumbarMR (94% sensitivity, 17% faster reporting), plus FDA 510(k) and CE for Prostate AI, an end-to-end prostate MRI workflow compatible with 11 fusion-biopsy systems and included in the UK TRANSFORM screening trial. Deployed across RadNet outpatient sites.

Conferences & voices

AIUM 2026 (Philadelphia, 27 to 29 May) put AI at the center of clinical and educational ultrasound, with Alison Noble (Oxford) giving a plenary on AI for ultrasound workflows and a pilot showing ChatGPT-5 matched instructor-led teaching for eFAST training. The first ACR-SIIM Practice Parameter for AI in imaging, approved at the ACR Annual Meeting, was framed as a landmark governance standard, with sites able to earn the ARCH-AI designation. Woojin Kim (HOPPR, ACR DSI) called automated report generation the hottest 2026 trend, while Stanford AIMI's Curtis Langlotz launched CheXthought, a globally representative reasoning dataset built through voluntary collaboration across 71 countries. A separate Radiology study (Schramm et al., TU Munich) found LLM assistance helped novice readers far more than expert neuroradiologists.

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.