Brain MRI AI achieves 97.5% accuracy across 52 diagnoses

RadAI Slice Newsletter

Weekly Updates in Radiology AI

Good morning, there. Michigan’s Prima AI reached 97.5% accuracy on 52 brain MRI diagnoses.

I’m encouraged by the sheer scale and breadth of what Prima achieved using over 220,000 MRI studies. This study stands out for its multiclass, real-world validation and offers a template for scalable neuroimaging AI. Prima’s potential for urgent triage and prioritization feels especially relevant as imaging demand and workforce pressures rise.

How would you integrate a multitask AI model like this into your own practice?

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

• Prima AI delivers near-instant, highly accurate brain MRI reads

• Medicare eyes coverage denial for AI brain MRI tools

• Foundation model BrainIAC generalizes across brain MRI tasks

• High-quality RCT: AI boosts mammography cancer detection, cuts intervals

• Plus: 3 newly released datasets, 6 FDA approved devices & 4 new papers.

🧠 Prima AI delivers near-instant, highly accurate brain MRI reads

RadAI Slice: A vision-language model trained on over 220,000 brain MRIs achieved up to 97.5% diagnostic accuracy.

The details:

• Tested on 30,000 MRIs spanning 52 neurologic disorders

• Mean AUC of 92.0% and robust urgency prioritization

• Outperformed prior models in neurodiagnostic accuracy

• Recommends subspecialty escalation for urgent findings

• Nature Biomedical Engineering publication; initial-stage evaluation

Key takeaway: Foundation models like Prima could enable scalable, rapid neuroimaging triage and diagnosis—particularly valuable as workloads grow and access inequities persist.

💡 Medicare eyes coverage denial for AI brain MRI tools

Image from: Radiology Business

RadAI Slice: A Medicare contractor signaled a possible denial of coverage for AI-based brain MRI analysis tools.

The details:

• National Government Services issued a draft non-coverage decision

• Impacts CPT codes 0865T/0866T across several US regions

• Cited lack of diverse datasets and clinical outcome evidence

• Denial could limit AI adoption for neurodegenerative imaging

• Public comment open through March 8

Key takeaway: Policy scrutiny remains high: robust validation and clinical utility data are prerequisites for reimbursement and AI clinical uptake in US imaging practices.

🧠 Foundation model BrainIAC generalizes across brain MRI tasks

Image from: EurekAlert

RadAI Slice: BrainIAC, a multicenter MRI foundation model, excelled at diverse clinical brain imaging tasks.

The details:

• Self-supervised pretraining on 49,000 brain MRIs

• Validated across seven clinical tasks—dementia, tumor, age, more

• Outperformed three task-specific models especially with limited labels

• NIH/NCI-funded, Nature Neuroscience publication

Key takeaway: Generalist imaging models like BrainIAC offer strong multi-task accuracy, potentially easing annotation barriers and supporting personalized neuroimaging care.

🩺 High-quality RCT: AI boosts mammography cancer detection, cuts intervals

Image from: Radiology Business

RadAI Slice: MASAI trial evidence shows AI-assisted mammography reduces interval cancers and improves sensitivity.

The details:

• Over 105,000 women; randomized, controlled Swedish trial

• AI cut future breast cancer diagnoses by 12% vs. standard

• AI trained globally on 200,000+ mammograms

• Lancet publication; led by Lund University

Key takeaway: RCT-level evidence supports cautious implementation of AI in screening, with clear patient benefit—but emphasizes the need for careful tool validation and ongoing monitoring.

BRISC (2026-01-27)

Modality: MRI | Focus: brain | Task: segmentation, classification

• Size: 6,000 scans, number of patients not specified

• Annotations: Expert-verified pixel-wise segmentation masks for tumors; class labels for glioma, meningioma, pituitary, and non-tumorous

• Institutions: Shahrood University of Technology, Iran University of Science and Technology (IUST), et al.

• Availability:

  public (Kaggle link)

• Highlight: Balanced multi-type tumor dataset with expert segmentations and multi-plane MRI images

MetaChest (2026-01-12)

Modality: X-ray | Focus: Chest, lung | Task: Multi-label classification, few-shot learning

• Size: 479,215 chest X-ray images from 4 datasets; covers 322,475 multi-labeled X-rays and 156,740 normal cases; ages 10-80

• Annotations: Multi-label classification for 15 common thoracic pathologies; derived from clinical reports (NLP) or expert radiologists. No segmentations.

• Institutions: Universidad Nacional Autónoma de México, Hospital San Juan et al.

• Availability:

  Public, code and partitions https://bereml.github.io/metachest/

• Highlight: Large-scale, multi-source X-ray dataset with few-shot and generalized few-shot learning partitions; enables realistic clinical episode generation.

K-MIMIC (2024)

Modality: X-ray, Bio-signals, EMR | Focus: ICU (Critical Care) | Task: Patient monitoring, Outcome prediction

• Size: 287,274 ICU admissions; 241,805 patients; 496,999 imaging studies; 22,588 bio-signal files

• Annotations: Structured EMR events, physiological waveforms, DICOM imaging—mainly chest X-ray; no manual segmentations

• Institutions: Seoul National University Hospital, The Catholic University of Korea, et al.

• Availability:

  Request-only via secure platform (link)

• Highlight: First multicenter multimodal Asian ICU dataset, linking EMR, waveforms, and imaging with temporal alignment.

🏛️ FDA Clearances

• K253057 - Siemens’ AI-Rad Companion Brain MR is cleared to support clinicians in automatic brain MRI segmentation and analysis, optimizing neuroimaging workflows.

• K253597 - Canon’s Aplio beyond and Aplio me V2.0 ultrasound system is cleared for high-quality Doppler imaging, aiding diagnostic confidence in clinical ultrasound.

• K253761 - HydroMARK Plus Breast Biopsy Site Marker (Dragonfly/Hummingbird shapes) is cleared to help accurately mark breast biopsy sites on imaging.

• K253735 - Philips’ AV Vascular solution is cleared for radiology image processing and vascular condition detection to support diagnosis and treatment planning.

• K253173 - United Imaging’s uCT 780 with dual-energy analysis is cleared for advanced CT imaging and detailed structural assessment.

• K250954 - Carestream’s DRX-Evolution Plus and DRX-Compass X-ray systems provide high-definition radiographic imaging, now FDA-cleared for clinical use.

• Explore last week's 7 radiology AI FDA approvals.

📄 Fresh Papers

• doi:10.1016/j.ijcard.2026.134221 - A multicenter prospective CMR study finds that integrating myocardial synchrony with strain improves CAD dysfunction detection (AUC up to 0.94) including in patients with preserved ejection fraction.

• doi:10.1038/s41551-025-01608-0 - University of Michigan’s Prima foundation model (trained on 220,000+ MRIs) achieved AUC 0.92 across 52 diagnoses and supports prioritization and differential diagnosis in brain MRI.

• doi:10.1038/s41593-026-02202-6 - BrainIAC, a 49,000-scan brain MRI foundation model, outperformed other pre-trained models on multiple neurodiagnostic tasks, excelling in data-scarce and high-difficulty settings.

• doi:10.1016/S0140-6736(25)02464-X - In the MASAI RCT (>105,000 women), AI-supported mammography reduced interval cancer rates and boosted sensitivity—without increasing recall rates—versus standard double reading.

• Browse 132 new radiology AI studies from last week.

📰 Everything else in Radiology AI last week

• A study found AI-assisted structured reporting boosted efficiency and accuracy in chest x-ray reporting compared to free-text workflows.

• Custom NLP ('FiNd') achieved over 96% accuracy for identifying ≥6mm lung nodules in chest CT reports, surpassing multiple general large language models.

• An XLNet-based open-source LLM achieved near 0.99 AUROC in automated CPT coding from IR postprocedure reports, cutting billing time and errors.

• Radiology Partners uses LLMs to extract findings, validate AI tool outputs, and support QA for imaging AI deployment assessment.

• Read 12 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.