Merlin AI foundation model interprets abdominal CT at scale

Good morning, there. Stanford’s Merlin 3D AI model achieved over 81% coding accuracy on 44,000+ abdominal CT scans.

I found it notable that Merlin was validated across thousands of external CT cases, showing strong accuracy in diagnostic coding and 5-year disease risk. This is a meaningful step towards foundation models that could assist radiologists in large-scale, multicenter CT interpretation and risk prediction, potentially easing workforce pressure and supporting new imaging biomarkers.

What role do you see for foundation models like Merlin in future CT reporting and workflow?

Here's what you need to know about Radiology AI last week:

• Merlin Foundation Model Advances Automated CT Interpretation

• RadNet’s $270M Gleamer Acquisition Expands Global Clinical AI Reach

• AI-Based Liquid Biopsy Detects Early, Reversible Liver Fibrosis

• Predeployment Evaluation Framework Guides Safe AI Model Adoption

• Plus: 3 FDA approved devices & 4 new papers.

🧠 Merlin Foundation Model Advances Automated CT Interpretation

RadAI Slice: The Merlin 3D vision-language model shows broad accuracy for abdominal CT coding, risk, and diagnosis.

The details:

• Over 81% accuracy for diagnostic coding across 692 ICD codes in external hospital data

• 90% accuracy for top 102 diagnoses, matching or exceeding specialist CT AI tools

• Accurately predicts 5-year chronic disease risk (75% AUROC)

• Generalizes to out-of-domain CT (e.g., chest), suggesting wide applicability

• Supports workflow efficiency, biomarker discovery, and is publicly available

Key takeaway: Validated AI foundation models like Merlin could raise the bar for automated CT interpretation, risk prediction, and scalable multicenter deployment, potentially accelerating both workflows and imaging science.

💰 RadNet’s $270M Gleamer Acquisition Expands Global Clinical AI Reach

Image from: Radiology Business

RadAI Slice: RadNet’s strategic purchase of Gleamer signals intensifying competition in clinical AI for radiology across multiple modalities.

The details:

• Gleamer serves 700+ customers in 44 countries, with 30M+ studies analyzed

• FDA-cleared AI spans X-ray, mammography, CT, MRI, and brain/lung/breast imaging

• Combined entity projects $30M in recurring revenue for 2024 and 90%+ yearly growth

• Positions DeepHealth as the sector’s dominant clinical AI provider

Key takeaway: This consolidation underscores rapidly rising demand for FDA-cleared, multi-modality AI—positioning large AI platforms as core partners for radiology groups seeking robust clinical solutions at scale.

🩺 AI-Based Liquid Biopsy Detects Early, Reversible Liver Fibrosis

Image from: EurekAlert

RadAI Slice: An AI-driven cfDNA fragmentome test identified early liver fibrosis in 1,576 patients—before cirrhosis is detectable by standard tests.

The details:

• AI recognized disease-specific cfDNA fragmentation patterns from WGS data

• High sensitivity for early-stage fibrosis; standard tests miss 50% of cirrhosis

• Assay also signaled risk for other chronic diseases (CV, inflammatory)

• Method validated in large, diverse chronic disease cohort

Key takeaway: This approach could enable routine noninvasive detection of reversible liver disease, supporting chronic disease management and expanding liquid biopsy’s radiological relevance.

🔬 Predeployment Evaluation Framework Guides Safe AI Model Adoption

Image from: Radiology Business

RadAI Slice: A practical new framework rates AI tasks by tediousness, risk of oversight, and clinical impact before purchase.

The details:

• Tested on 13 Aidoc models, 88,000 exams, across multiple sites

• Task-level value: 5 models high, 5 medium, 2 low

• Framework aligns inherent value with actual clinical needs

• Validated via post-deployment radiologist surveys

Key takeaway: Structured predeployment vetting can bridge the gap between AI marketing and real-world outcomes, improving investment and patient care in radiology.

🏛️ FDA Clearances

• K252500 - CARA System receives 510(k) clearance for real-time interventional fluoroscopy X-ray imaging in radiology procedures.

• K260205 - AS Software Version Asera cleared as an image processing platform for diagnosis and planning in radiology.

• K254018 - Portable Dental X-ray Device (GT-1) cleared for mobile/point-of-care dental imaging, facilitating broad access.

📄 Fresh Papers

• doi:10.1016/j.compmedimag.2026.102740 - DiffusionTBAD used text-guided diffusion to generate synthetic aortic dissection CTAs, improving model accuracy and segmentation with few real cases.

• doi:10.64898/2026.02.26.26347229 - AI trained/externally validated on 50,000 echocardiograms achieves AUC 0.92 for MVP and 0.88 for significant MR—across UCSF and Houston Methodist.

• doi:10.1038/s41586-026-10181-8 - Nature: Merlin 3D vision-language CT model trained on 6 million slices, 1.8M codes; open dataset/models released for broad benchmarking.

• doi:10.1177/13524585261421491 - Multi-modal deep network in prospective study automates MR classification of paramagnetic rim/remyelinated lesions in 180 MS patients.

• Browse 147 new radiology AI studies from last week.

📰 Everything else in Radiology AI last week

• Harrison.ai seeks FDA exemption for some CAD AI devices, triggering debate over balancing fast access with safety in radiology AI.

• European radiologists at ECR 2026 call for more and better evidence as new EU HTA rules reshape AI reimbursement and clinical expectations.

• U-Net automates chest CT muscle mass estimation in 1,235 CanCOLD patients, improving COPD sarcopenia detection and prognostication.

• AI creates virtual immunostains on label-free thyroid tissues, improving pathologic efficiency and diagnostic workflow.

• Ultrasound AI’s Delivery Date AI wins FDA De Novo clearance, reaching 0.92 R² on 5,700+ prenatal exams for real-time delivery date prediction.

• Read 18 fresh radiology AI news articles for last week.

That's it for today!

👋 Quick favor: drag this into your Primary tab so you don’t miss next week. Or just hit Reply with one thought.

See you next week.

P.S. We keep building free tools to accelerate your radiology work. What's the most time-consuming pain point in your day that we should help speed up? Reply and share your take so we keep building around you.