🩻 0.913 AUC for generalist abdominal CT AI

Good morning, there. RADAR achieved mean AUC 0.913 across 146 abdominal CT findings.

I see this as a meaningful test of generalist imaging AI beyond single finding tools. External AUC 0.895 and about a 10% sensitivity gain for readers make it more relevant than a benchmark only result.

How would you validate a generalist CT model before routine use?

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

• 🩻 RADAR tests generalist AI across 146 CT findings

• CTA aneurysm AI shows real workflow yield

• Open source registration adapts in 5 minutes

• ESUR sets out guardrails for prostate MRI AI

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

🩻 RADAR tests generalist AI across 146 CT findings

RadAISlice: This study stood out to me for its scale, external testing, and direct reader impact.

The details:

• Trained on 424,911 contrast enhanced abdominal CT exams

• Used 15M anatomy specific image text pairs

• Mean AUC was 0.913 across 146 CT findings

• External testing across 8 centers held AUC at 0.895

• Reader sensitivity improved by about 10% with assistance

Key takeaway: This moves generalist CT AI closer to practical decision support, but local validation and workflow design will still determine how useful it is.

🧠 CTA aneurysm AI shows real workflow yield

RadAISlice: This study stands out because it measured incremental yield, not just standalone accuracy.

The details:

• 3,856 consecutive brain CTAs ran in shadow mode

• AI sensitivity was 0.846 versus 0.718 for radiologists

• Concordance between AI and reports was 96.3%

• AI added 55 true positives and 46 false positives

• Most AI only aneurysms were under 3 mm

Key takeaway: The result supports targeted CTA AI deployment, especially in acute settings where added detections appear most efficient.

🧭 Open source registration adapts in 5 minutes

RadAISlice: This Nature paper feels relevant to image guided procedures because it reduces the need for anatomy specific retraining.

The details:

• Uses each patient's preoperative CT or MRI for simulation

• Adapts to new anatomy with about 5 minutes of fine tuning

• Performs registration in seconds

• Evaluated on real fluoroscopy across anatomy and hospitals

• Open source software is available for research use

Key takeaway: Fast patient specific registration could make navigation less procedure specific if it proves robust in clinical deployment.

🧩 ESUR sets out guardrails for prostate MRI AI

RadAISlice: I like this paper because it focuses less on model performance and more on what responsible deployment actually requires.

The details:

• Recommends keeping radiologists in the loop

• Emphasizes local and prospective validation before routine use

• Calls for user training around AI limitations

• Highlights ongoing quality assurance and drift monitoring

Key takeaway: The harder part of prostate MRI AI may be maintaining safe performance after deployment, not simply reaching a strong benchmark.

🏛️ FDA Clearances

• K262060 - 3D RD S cleared as radiological image processing software for enhanced image analysis and planning.

• K260525 - NextMR Spine cleared to automate processing of spine MR images for radiology analysis workflows.

• K260488 - Siemens MULTIX Impact added AI based acquisition features for thorax and long leg or spine exams.

• K261718 - Radformation EZFluence cleared for fluence optimization in external beam radiotherapy planning.

📄 Fresh Papers

• doi:10.2214/AJR.26.35514 - An expert panel backed radiologist in the loop AI for ARIA detection with quality assurance and prospective monitoring.

• doi:10.1148/rg.260009 - Radiographics reviewed opportunistic CT screening, including deployment, quality assurance, and actionability limits.

• doi:10.1148/radiol.260238 - MESA data linked CT paraspinal muscle attenuation to later thoracic BMD loss and vertebral fracture risk.

• doi:10.1038/s41591-026-04488-2 - I3LUNG tested explainable AI for NSCLC immunotherapy prediction in 2,396 patients with multimodal data.

• Browse 283 new radiology AI studies from last week.

📰 Everything else in Radiology AI last week

• DeepHealth received FDA 510(k) clearance for chest X ray AI built on a foundation model to detect and localize thoracic abnormalities.

• Radiology leaders urged departments to define clinical problems, workflow fit, and monitoring plans before buying AI.

• Cognita Imaging received a $1.29M FDA research contract to test LLM as jury methods for radiology report evaluation.

• A NEJM perspective argued that AI may expand clinical work by increasing demand and creating new oversight roles.

• Read 11 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.