Weekly AI digest

Three things that mattered

• Multimodal LLMs still fall short of clinical usefulness

  RadM-Bench, a bilingual 720-case benchmark across 9 subspecialties, scored 10 multimodal LLMs below 1.5 of 3 on both teaching and routine clinical cases. Radiologist-selected 2D key images improved every model by 19.8% to 139.2%, yet volumetric input degraded all evaluable models and performance on curated teaching cases did not transfer to clinical routine.

• Dataset diversity beats dataset size in segmentation bias

  An AJNR audit of four glioblastoma MRI segmentation models on 480 independent patients found the model trained only on white non-Hispanic men scored lowest (Dice 0.943 FLAIR), while the moderate but diverse BraTS 2024 model scored highest (0.996) and outperformed the far larger FeTS model. Demographic heterogeneity reduced bias even without more data.

• A transfusion model that quantifies its own uncertainty

  The MCGB framework predicted transfusion need in 849 upper GI bleeding patients with AUROC 0.97 and sensitivity 0.99 on a fully held-out hospital, and estimated dose with conformal 95% prediction intervals covering 0.94. Clinical monotonic constraints and cross-site testing address the standard objections to bedside machine learning.

Peer-reviewed

Selected from 207 papers indexed in PubMed this week.

• Prehospital Injury Severity Estimate (PHISE) matches in-hospital trauma scores when embedded in AI models

  NPJ Digital Medicine, 7 August 2026

• Prediction of Blood Transfusion Need and Dose in Patients With Upper Gastrointestinal Bleeding: Retrospective Multicenter Prediction Model Study

  JMIR Medical Informatics, 4 August 2026

• Evaluating Sociodemographic Biases in Artificial Intelligence-Based Glioblastoma Response Assessment Algorithms

  AJNR American Journal of Neuroradiology, 3 August 2026

• Machine Learning to Identify Point-of-Care Ultrasound and Evaluate Standardized Documentation: Retrospective Operational Cohort Study

  Journal of Medical Internet Research, 7 August 2026

• A Bilingual Benchmark for Evaluating Diagnostic Performance of Multimodal Large Language Models in Radiology (RadM-Bench): Evaluation Development and Validation

  Journal of Medical Internet Research, 7 August 2026

• Large Language Models for Classifying Usual Interstitial Pneumonia from Radiology Reports: Native Reasoning Versus Structured Prompting

  Journal of Imaging Informatics in Medicine, 5 August 2026

• Automatic Patient Eligibility for Photon-counting CT Using Discriminative and Generative AI Models in Neuroradiology

  Academic Radiology, 5 August 2026

• Human-AI Interaction With AI-Assisted Tumor Overlays in Pediatric Whole-Body Magnetic Resonance Imaging: Exploratory Reader Study

  JMIR Human Factors, 7 August 2026

Industry & regulation

• GE HealthCare launches Invenia ABUS Prime and StreamVue

  GE HealthCare launched its Invenia ABUS Prime automated breast ultrasound system alongside ABUS StreamVue, a browser-based enterprise viewer that recently received FDA 510(k) clearance and CE marking for dense-breast supplemental screening programs.

• Cortechs.ai expands EU MDR CE marking

  Cortechs.ai announced expanded CE marking under EU MDR for Version 5 of NeuroQuant, NeuroQuant Brain Tumor and OnQ Prostate, enabling commercialization of its current quantitative imaging platform across the European Economic Area.

• 4DMedical takes global distribution of RevealAI-Lung

  4DMedical announced a $3.4M strategic investment in RevealDx and a distribution agreement covering the US, Europe, Australia and New Zealand for RevealAI-Lung, an FDA-cleared nodule characterization tool it will embed in its contextflow chest CT platform.

• GE HealthCare introduces LOGIQ e compact ultrasound family

  GE HealthCare introduced the LOGIQ e Xi and Si laptop-format ultrasound systems, combining console-level performance with AI features such as Nerveblox nerve-block landmark labeling and automated cardiac function imaging.

Conferences & voices

RSNA issued the official launch of its 2026 Knee Abnormality Detection AI Challenge, the first of its AI challenges to focus on musculoskeletal MRI and the first to combine images with radiology report text, drawing on more than 5,000 knee MRI exams with reports in 12 languages from 16 sites on five continents. The Kaggle competition runs through 22 October with $77,000 in prizes and winners recognized at RSNA 2026 in Chicago.

Curt Langlotz amplified the publication of MedVAL in npj Digital Medicine, a Stanford paper he coauthored proposing language models as scalable validators of AI-generated medical text, and revisited his 2017 line about radiologists who use AI replacing those who do not after meeting the Raidium team in Palo Alto.

Woojin Kim announced he will represent the ACR Data Science Institute at EuSoMII 2026 in Heraklion this October, after teaching entrepreneurship and AI in radiology in Geraldine McGinty's healthcare leadership course at Weill Cornell earlier in the week.

Amine Korchi summarized a newly published systematic review with the Geneva University Hospitals neuroradiology team on vision-language foundation models for 3D neuroradiological interpretation: domain-specific models consistently beat general-purpose ones and bring AI-assisted report drafting closer to clinical reality, but too many required corrections, heavy compute, hallucinations and regulatory gaps keep them out of routine practice for now.

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.