Weekly AI digest

Three things that mattered

• Pathology foundation model learns from report dialogue

  PRISM2 (Vorontsov E et al., Nature Medicine) was trained on 2.3 million whole-slide images and 14 million question-answer pairs derived from 700,000 pathology reports, using clinical dialogue rather than labels alone as the supervision signal. With prompt-based inference it matches or exceeds clinical-grade products for cancer detection in the prostate, breast and breast lymph node (P < 0.05), though the abstract publishes comparative claims without the underlying numbers.

• Reproducibility splits two chatbots reading thyroid reports

  Xie Y et al. (Journal of Medical Internet Research) tested ChatGPT-4o and DeepSeek-R1 on 1,063 thyroid ultrasound reports from 3 centres. DeepSeek-R1 was more sensitive (0.879 versus 0.692) and more accurate (0.729 versus 0.644), but the decisive gap was stability on benign versus malignant calls, kappa 0.869 versus 0.609, with both models remaining below senior radiologists at AUC 0.865.

• Discarded CT attenuation maps carry prognostic signal

  Zhou J et al. (European Journal of Nuclear Medicine and Molecular Imaging) applied AI to the CT attenuation maps acquired alongside myocardial perfusion imaging in 4,552 patients drawn from 43,099 studies across 15 sites, all with mild calcium and normal perfusion. Proximal calcium, present in 38%, independently predicted major adverse cardiovascular events (adjusted hazard ratio 1.24) and death (adjusted hazard ratio 1.25) over 3.6 years, adding 12% net reclassification beyond the calcium score.

Peer-reviewed

Selected from 232 papers indexed in PubMed this week.

• End-to-end multimodal pathology foundation model with clinical dialogue

  Nature Medicine, 31 July 2026

• The Performance of ChatGPT-4o and DeepSeek-R1 in Interpreting Thyroid Nodule Ultrasound Text Reports: Multicenter Study

  Journal of Medical Internet Research, 28 July 2026

• Multi-modal Stacking Machine Learning Model for Predicting Lymph Node Metastasis in Cervical Cancer

  Academic Radiology, 29 July 2026

• Impact of Large Language Model-Based AI Tools on Physician-Patient Communication: Systematic Review and Meta-Analysis

  Journal of Medical Internet Research, 31 July 2026

• Hidden risk in normal myocardial perfusion scans: AI-detected proximal coronary calcium on CT attenuation maps improves prognosis

  European Journal of Nuclear Medicine and Molecular Imaging, 28 July 2026

• Systematic Review of AI-Driven Applications for Screening Nasopharyngeal Carcinoma Using MRI

  Journal of Magnetic Resonance Imaging, 26 July 2026

• Social Media Discourse on Breast Cancer Screening Barriers in Japanese and English: Cross-Sectional Infodemiology Study

  Journal of Medical Internet Research, 31 July 2026

• Bridging the AI chasm in oncology: a standardized platform to enable the in silico clinical validation of AI models within the CHAIMELEON Project

  European Radiology Experimental, 31 July 2026

Industry & regulation

• DeepHealth cleared for AI breast ultrasound

  RadNet subsidiary DeepHealth received FDA 510(k) clearance for DeepHealth Breast Ultrasound, validated in a multi-reader multi-case study with 16 US board-certified radiologists and slated for deployment across RadNet's network by the end of 2026.

• ThinkSono cleared for AI ultrasound guidance in DVT

  ThinkSono Guidance received FDA 510(k) clearance as the first AI-powered ultrasound guidance software cleared for evaluation of suspected deep vein thrombosis, guiding non-ultrasound-trained staff who completed 60 to 90 minutes of standardized training.

• Caristo wins De Novo authorization for coronary inflammation

  The FDA granted De Novo authorization for CaRi-Heart, the first technology authorized in the US to quantify coronary inflammation from routine coronary CT angiography and generate a 10-year cardiovascular mortality risk estimate.

• Median and Olea Medical partner on lung screening AI

  Median Technologies signed a global commercial agreement with Canon Medical Systems subsidiary Olea Medical to deploy the lung cancer screening software eyonis LCS across the United States and Europe, following FDA clearance in February and CE marking in July 2026.

• Vanderbilt and Siemens sign $87 million imaging partnership

  Vanderbilt Health and Siemens Healthineers launched a multi-year, $87 million partnership covering MRI, CT, molecular imaging, interventional radiology and radiation oncology systems, with joint work planned on data infrastructure and AI-enabled workflow innovation.

Conferences & voices

The European Society of Radiology published the report from its European Parliament event on imaging and cardiovascular health, co-organised with the European Association of Nuclear Medicine, naming responsible integration of artificial intelligence as a cross-cutting priority for delivering the EU Safe Hearts Plan. The European Board of Radiology opened registration for the European Diploma in Radiology at ECR 2027 and launched the EDiR Academy, consolidating its preparation resources into a single structured pathway.

Coverage published this week from the Royal Australian and New Zealand College of Radiologists meeting held in Sydney on 24 and 25 July, which drew more than 250 professionals, framed the practical AI agenda as education, data maturity, guard rails and patient consent rather than workforce displacement. Separately, The Imaging Wire reported a Radiology study in which an AI safeguard review across 109 facilities and 578,000 screening exams lifted the cancer detection rate of general radiologists by 25%, closing the gap with specialists.

Curt Langlotz posted twice on 30 July, first on the Circulation editorial he co-wrote with Fatima Rodriguez arguing that the time for clinical implementation of coronary calcium detection on non-gated CT is now, and second on a recorded conversation about AI, the radiologist workforce and patient consent. Pranav Rajpurkar highlighted the ReXGrounding CT challenge at MICCAI 2026, where twelve teams grounding free-text findings to 3D CT segmentations are clustered near a mean Dice of 0.33, and Daniel Pinto dos Santos shared a EuSoMII briefing on AI in musculoskeletal radiology.

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.