11089 MRIs train AI to flag tumor enhancement

Good morning, there. A multicohort model used 11089 brain MRIs to detect tumor enhancement with 83.0 percent balanced accuracy.

I see this as a practical gadolinium stewardship signal, not a replacement for contrast MRI. It could help radiology teams decide when a planned noncontrast exam needs contrast added. Pediatric performance still needs work.

Could this change when you add gadolinium during brain tumor MRI?

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

• Noncontrast brain MRI predicts enhancement

• Prostate MRI AI reduces biopsy burden

• Generative AI still trails radiologists

• RSNA releases aneurysm AI dataset

• Plus: 2 newly released datasets, 6 FDA approved devices & 4 new papers.

🧠 Noncontrast brain MRI predicts enhancement

RadAISlice: I am watching this as a realistic triage use case for gadolinium sparing neuro oncology MRI.

The details:

• 11,089 brain MRI studies came from 10 datasets in 4 countries

• nnU Net reached 83.0 percent balanced accuracy for enhancement detection

• Sensitivity was 91.5 percent and specificity was 74.4 percent

• Radiologist majority vote reached 71.7 percent balanced accuracy in 564 cases

• Only 45 percent of pediatric cases reached Dice of at least 0.3

Key takeaway: This could support contrast add on decisions for brain tumor MRI, but it should guide protocol triage rather than replace postcontrast imaging.

🧬 Prostate MRI AI reduces biopsy burden

Image from: Radiology Business

RadAISlice: This result feels clinically useful because it targets the gray zone around prostate MRI biopsy decisions.

The details:

• AI was compared with radiologist assigned PI RADS on multiparametric MRI

• The approach avoided about 19 percent of prostate biopsies

• Sensitivity for clinically significant cancer remained 98.4 percent

• Findings were published in European Radiology

• Biopsy decisions still remain clinical and multidisciplinary

Key takeaway: A tool like this could reduce low value biopsies after prostate MRI while keeping radiologists central to risk communication.

📊 Generative AI still trails radiologists

RadAISlice: This meta analysis adds useful caution to current enthusiasm for image enabled generative AI.

The details:

• 48 studies were included in the PRISMA DTA review

• Pooled accuracy was 42.9 percent for free text tasks

• Pooled accuracy was 58.1 percent for choice tasks

• Expert physicians were 13.0 percentage points more accurate overall

• Text only inputs outperformed image only inputs by 25.9 points

Key takeaway: For radiology, the near term role looks more like supervised text support than autonomous diagnostic image interpretation.

🧠 RSNA releases aneurysm AI dataset

RadAISlice: This dataset stands out because aneurysm AI needs diverse angiographic and routine neuroimaging examples.

The details:

• 7,202 CTA, MRA, and MRI series came from 4,278 adult patients

• Data came from 21 institutions across 12 countries and 5 continents

• 5,041 series are publicly available for noncommercial research

• 178 series include 3D segmentations of vascular locations

• The dataset powered the 2025 RSNA aneurysm detection challenge

Key takeaway: Better public benchmarks should make aneurysm AI claims easier to compare across modalities, sites, and lesion types.

DrugTargetWorld (Unspecified)

Modality: MRI | Focus: Heart, cardiovascular disease | Task: Drug target discovery, causal inference

• Size: 20 synthetic worlds. Each has 54,000 participants and 10,800 cardiac MRI studies.

• Annotations: Hidden causal ground truth. Labels include driver proteins, bias decoys, intervention direction, and phenotype scoring targets. No released segmentations.

• Institutions: Stanford University, Brown University

• Availability:

  Public: GitHub, Hugging Face

• Highlight: Procedural synthetic biobank with concealed causal models for verifiable AI scientist benchmarking.

RSNA-ICA (2025 (exact date unspecified))

Modality: CTA, MRA, MRI | Focus: Brain, intracranial arteries | Task: Aneurysm detection, localization

• Size: 7,202 image series from 4,278 adults. Public MIRA release has 5,041 series.

• Annotations: Expert labels for aneurysm presence and 13 vascular locations. Point localizers; 3D vessel segmentations for 178 series.

• Institutions: RSNA, ASNR, et al.

• Availability:

  Public for noncommercial use: MIRA. Challenge data also on Kaggle.

• Highlight: Largest public expert-annotated intracranial aneurysm dataset. Includes CTA, MRA, and routine MRI from 21 institutions.

🏛️ FDA Clearances

• K260055 - Siemens syngo.CT Brain Quantification segments CT hyperdensity and measures midline shift on noncontrast head CT.

• K260023 - SimBioSys TumorSight Risk estimates breast cancer recurrence risk using DCE MRI plus clinical pathology data.

• K262390 - BoneXpert estimates pediatric bone age from hand radiographs for patients ages 2 to 21 years.

• K261368 - Rapid Platform R6.7 supports automated CT, CTA, CTP, CBCT perfusion, and MRI analysis for acute stroke.

• K262548 - Philips EPIQ Inspira and Affiniti add AI assisted left atrial volume and strain quantification for echo.

• K261125 - Samsung NeuroLogica BodyTom 64 and OmniTom Elite add CT reconstruction software that reduces noise and artifacts.

• Explore last week's 12 radiology AI FDA approvals.

📄 Fresh Papers

• doi:10.1186/s41747-026-00791-2 - LIBRA used 1,026 patients from 7 UK hospitals to build lung nodule radiomics models with AUC 0.91.

• doi:10.1093/ehjci/jeag270 - AI quantitative CCTA improved ACS culprit lesion prediction in both women and men in 351 EMERALD II patients.

• doi:10.1148/radiol.260703 - In 313 mild IPF patients, deep learning CT vascular metrics linked venous remodeling to acute exacerbation risk.

• doi:10.1016/j.jocmr.2026.102806 - Prospective CMR data found one heartbeat free breathing deep learning cine comparable for mitral regurgitation.

• Browse 225 new radiology AI studies from last week.

📰 Everything else in Radiology AI last week

• Everlywell is selling FDA authorized Clairity Breast risk prediction to consumers for 249 dollars per upload.

• MRI2CT gained FDA clearance for NextMR, which creates 3D bone views from routine spine MRI in under 60 seconds.

• The EU CHAIN project will spend 20 million euros to test validated cardiovascular AI across 13 countries.

• AIRS Medical won FDA clearance for MRI body composition analysis with automated reports in under 5 minutes.

• A 47 study review urges life cycle reporting as imaging AI contributes to rising data center energy demand.

• 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.