94.4% accuracy for contrast free scar MRI

Good morning, there. VNE identified myocardial infarction with 94.4% accuracy in confident cases.

I see this as a practical test of whether AI can reduce gadolinium use without weakening scar assessment. The result matters for CMR access, renal risk, and protocol design.

Would you let AI decide when gadolinium is still needed?

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

• Contrast free scar MRI gets multicenter test

• Mammography AI changed how readers looked

• EU AI Act pushes surveillance in radiology AI

• MRI foundation model spans 44 clinical tasks

• Plus: 5 newly released datasets, 6 FDA approved devices & 4 new papers.

🔎 Radiologist seeking remote teleradiology work

A radiologist licensed in Italy and Switzerland is looking for remote teleradiology opportunities.

Hiring or know someone who is? Reply to this email and I’ll make an introduction.

🫀 Contrast free scar MRI gets multicenter test

RadAI Slice: This prospective CMR study feels close to daily protocol decisions.

The details:

• VNE used cine and T1 maps only, with LGE as ground truth

• MI diagnosis reached 94.4% accuracy in 107 confident cases

• Across all 136 cases, accuracy was 87.5%

• VNE scar quantification correlated with LGE at R 0.90

• Readers would skip LGE in 69.7% of referrals

Key takeaway: This could make CMR scar imaging more selective, with AI helping decide when gadolinium still adds value.

👀 Mammography AI changed how readers looked

RadAI Slice: This Radiology reader study is a useful reminder that prompts reshape search behavior.

The details:

• 10 NHS readers reviewed 60 screening mammography cases

• False negative AI reduced median sensitivity from 71% to 39%

• False positive AI raised specificity from 21% to 39%

• Visible prompts increased median read time from 25s to 34s

• Eye tracking showed fewer fixations when AI missed cancers

Key takeaway: For breast AI adoption, false negative behavior and prompt design may matter as much as headline AUC.

🇪🇺 EU AI Act pushes surveillance in radiology AI

RadAI Slice: This opinion paper usefully translates regulation into imaging AI operations.

The details:

• Article 4 requires AI literacy for providers and deployers

• Article 72 requires lifecycle monitoring for high risk AI

• Medical AI post market rules are slated for 2 August 2028

• Authors call for drift checks, subgroup audits, and shutdown criteria

Key takeaway: Radiology AI governance is moving toward local competence, audit trails, and live performance surveillance.

🧠 MRI foundation model spans 44 clinical tasks

RadAI Slice: This foundation model paper stands out for breadth across real MRI use cases.

The details:

• Pretraining used 336476 volumetric MRI scans from 34 datasets

• Corpus covered 10 anatomical structures and multiple sequences

• Benchmark included 44 tasks across diagnosis, segmentation, and reports

• MARS ranked first in 41 of 44 benchmarks

• External datasets tested heterogeneous real world performance

Key takeaway: For radiology AI teams, multi sequence pretraining may help reduce the one model per task problem.

MedRealMM (2026-07-16)

Modality: Multimodal images + text | Focus: Multisystem, skin | Task: Next-response generation, clinical response evaluation

• Size: 5,620 consultation cases; patients not specified; 1–10 uploaded images per case.

• Annotations: MCCP labels, patient intent/stage labels, and physician-refined case-specific rubrics with positive and negative criteria.

• Institutions: JD Health International Inc., Shanghai Artificial Intelligence Laboratory, et al.

• Availability:

  restricted via DUA; Hugging Face

• Highlight: Real Chinese online consultations with patient-uploaded images and physician-in-the-loop rubrics.

SYMBALBENCH (2026-07-16)

Modality: X-ray, natural images | Focus: Chest, common objects | Task: Systematic misalignment detection, caption auditing

• Size: 1.7M image-text pairs across 420 datasets. Base data include 4,349 COCO images and 2,233 CXR scans. Patients not specified.

• Annotations: Ground-truth systematic misalignment labels. Each label pairs an erroneous textual fact with an associated visual feature. Reference captions are included in a variant.

• Institutions: Stanford University, HOPPR

• Availability:

  Restricted + public code: GitHub

• Highlight: First benchmark for detecting systematic image-text caption errors, including medical CXR settings.

ERDES (2026)

Modality: US (POCUS) | Focus: Eye, retina | Task: RD classification, macular status classification

• Size: 5,381 ocular US video clips. Patient count not specified; multiple clips may come from one encounter.

• Annotations: Clip-level diagnostic labels. Non-RD/RD, Normal/PVD, and macula-intact/macula-detached. No frame-level annotations.

• Institutions: The Ohio State University; University of Arizona

• Availability:

  Public: Zenodo

• Highlight: First open ocular POCUS video dataset with both RD and macular status labels.

Cet.CT-Bank (15 July 2026)

Modality: CT | Focus: Whole body, head | Task: Image analysis, classification

• Size: 8 PMCT cases from 8 cetaceans. 26,589 axial slices and 26,993 DICOM images.

• Annotations: No manual segmentations. Includes DICOM metadata, validation reports, species/biological data, and necropsy etiologic diagnoses.

• Institutions: University of Las Palmas de Gran Canaria, IDeTIC

• Availability:

  Public: Zenodo

• Highlight: Open standardized PMCT DICOM bank for stranded cetaceans, with complete imaging metadata and necropsy context.

SWM Atlas (July 6, 2026)

Modality: dMRI | Focus: Brain; superficial white matter | Task: Tract segmentation; connectivity mapping

• Size: 171 7T dMRI scans from 171 healthy HCP participants. Atlas has 643 Yeo-7 and 1,403 Yeo-17 clusters.

• Annotations: SWM tractograms, cluster labels, Yeo 7/17-network labels, Neurosynth terms, and TW-dFC uncertainty maps.

• Institutions: Nanjing University of Science and Technology; Tsinghua University; et al.

• Availability:

  Public: Science DataBank

• Highlight: Fine-grained 7T dMRI SWM atlas with network organization and functional annotations.

🏛️ FDA Clearances

• K260528 - OncoStudio cleared for CT based radiation therapy planning support and tumor targeting workflow.

• K260576 - MeVis AVM Contour cleared for 2D DSA to 3D CT registration in cranial radiosurgery planning.

• K261119 - NorthStar Mapping System cleared to support real time MR guided cardiovascular procedures.

• K253628 - Auto Seg Spine cleared for AI segmentation and labeling of spine and pelvis CT for surgical planning.

• K253969 - ADAS 3D cleared for cardiac CT and MR segmentation, scar visualization, and EP procedure planning.

• K260667 - Philips Alturion cleared with AI Auto Measure Abdomen for automated ultrasound measurements.

• Explore last week's 11 radiology AI FDA approvals.

📄 Fresh Papers

• doi:10.1038/s41698-026-01607-7 - DeepCT MTM predicted aggressive HCC subtype across 3118 patients from 20 hospitals, with AUC 0.845 externally.

• doi:10.1016/j.acra.2026.06.037 - Prospective 16 institution ultrasound model identified early renal fibrosis with external AUC 0.823.

• doi:10.1007/s00330-026-12745-8 - A European Radiology review found lung nodule AI evidence is growing, but outcome and cost data remain limited.

• doi:10.1097/RLI.0000000000001303 - Chest radiograph AI detected osteoporosis with AUC 0.94 and linked scores to later fracture and mortality.

• Browse 371 new radiology AI studies from last week.

📰 Everything else in Radiology AI last week

• Radiologists identified AI generated images correctly 75% of the time, with CT and MRI fakes hardest to spot.

• Radiology represented 76.5% of 1430 FDA authorized AI and ML devices in a longitudinal review.

• A Penn State study found patient confidence dropped when AI disagreed with a clinician assessment.

• A hardware random number generator embedded image signatures that may help detect medical image tampering.

• Read 9 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.