🩻 Mammography AI could cut second reads by 77%

Good morning, there. AI marked 76.6% of negative screening mammograms as low risk in a 55,589-exam study.

This is one of the clearest workflow stories of the week. Instead of asking whether AI can replace a reader, the study asks a more practical question: can AI safely redesign part of the reading pathway? The result is promising, but the 1 cancer in the low-risk group also makes the governance question impossible to ignore.

How much missed cancer risk is acceptable when AI reduces screening workload?

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

• 🩻 AI triage targets second reads in mammography

• Patient safety frame for radiology AI

• MRI muscle marker flags frailty in pediatric brain tumors

• OpenRad maps open radiology AI models

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

🩻 AI triage targets second reads in mammography

RadAI Slice: This study stands out because it tests AI against a real screening bottleneck.

The details:

• 55,589 BI-RADS 1 or 2 mammograms from 42,419 women were analyzed

• Second readers recalled 183 exams and detected 12 cancers

• AI labeled 42,606 exams as low risk, equal to 76.6% of the cohort

• The AI low-risk group included 1 cancer and had 0.47 interval cancers per 1,000 exams

• Bypassing AI low-risk exams could reduce second reading workload by about 77%

Key takeaway: This could move radiology AI from replacement narratives toward risk-based worklist design, but any low-risk bypass strategy needs prospective monitoring and clear governance.

🛡️ Patient safety frame for radiology AI

RadAI Slice: This report is a useful companion to the mammography triage study because it focuses on what happens after AI enters clinical workflow.

The details:

• Radiology AI Special Report centers policy around patient safety

• It covers data stewardship, validation, deployment, and monitoring

• The framework connects regulatory science with implementation science

• Shared responsibilities span vendors, institutions, clinicians, and policymakers

Key takeaway: Radiology groups need more than clearance status. They need lifecycle oversight before and after AI reaches PACS.

🧠 MRI muscle marker flags frailty in pediatric brain tumors

RadAI Slice: This study shows how routine surveillance MRI can become a longitudinal marker of patient condition, not only tumor status.

The details:

• 5,661 MRIs from 881 children and survivors across 3 cohorts

• iTMT sarcopenia appeared in 531 of 730 patients with weight data

• Sarcopenic overweight appeared in 215 of 730 patients

• High-grade glioma survival was worse with sarcopenia at diagnosis, p = 0.046

Key takeaway: This could expand the role of radiology AI from lesion assessment to supportive care signals during long-term follow-up.

📚 OpenRad maps open radiology AI models

RadAI Slice: This repository is practical because radiology AI model discovery remains scattered across papers, GitHub repositories, demos, and datasets.

The details:

• 5,239 works were screened for open radiology AI models

• 1,694 model records span CT, MRI, X-ray, ultrasound, and more

• 10 expert reviewers manually verified LLM-generated metadata

• MRI led the repository, including 621 neuroradiology AI models

Key takeaway: OpenRad could make evaluation, benchmarking, and local experimentation easier by making code, weights, demos, and model metadata easier to find.

OpenRad (28 May 2026)

Modality: CT, MRI, X-ray, US | Focus: Multi-organ; neuroradiology and chest prominent | Task: Model discovery; benchmarking/reuse support

• Size: 1,694 open-access radiology AI model records. No patient or scan data released.

• Annotations: Standardized model records. Includes modality, subspecialty, use case, architecture, metrics, code links, weights, and demo availability.

• Institutions: University of Crete; Karolinska Institute; et al.

• Availability:

  Public: OpenRad

• Highlight: Large curated index of open-access radiology AI models with verified code, weights, and demo metadata.

OMAMA-DB (2024)

Modality: MG/DBT | Focus: Breast | Task: Cancer classification; lesion detection/localization

• Size: 231,080 images; patient count not reported. Includes 163,568 2D mammograms and 67,512 3D DBT volumes.

• Annotations: Pathology-based cancer labels. DeepSight automated lesion bounding boxes with confidence scores for 7,725 cancer cases.

• Institutions: University of Massachusetts Boston, DeepHealth/RadNet, et al.

• Availability:

  Public: Harvard Dataverse

• Highlight: Large public breast screening dataset combining 2D mammograms and 3D DBT with pathology labels and automated lesion boxes.

PHLF (2025-12-09)

Modality: MRI | Focus: Liver; hepatobiliary system | Task: Segmentation; PHLF prediction

• Size: 220 patients with 220 HBP Gd-EOB-DTPA MRI scans. 14,895 slices from 3 centers.

• Annotations: Expert masks for liver, Couinaud segments, liver tumors, spleen, and psoas muscle. Includes clinicopathological variables and ISGLS grade B/C PHLF outcomes.

• Institutions: Peking University Shenzhen Hospital; First Affiliated Hospital of Shantou University Medical College; et al.

• Availability:

  Public: Zenodo

• Highlight: First multi-center Gd-EOB-DTPA liver MRI dataset for combined FLR volume and function assessment with PHLF outcomes.

TrackRAD2025 (March 15, 2025)

Modality: MRI | Focus: Thorax, abdominopelvic | Task: Target tracking, segmentation

• Size: 585 patients. 2D+t sagittal cine-MRI with more than 2.878M unlabeled frames. 50 labeled and 477 unlabeled training cases; 58 private test cases.

• Annotations: Expert target masks for 108 patients. More than 10,000 labeled frames, including more than 8,000 multi-observer labels.

• Institutions: LMU University Hospital, Sichuan Cancer Hospital & Institute, et al.

• Availability:

  Public training data; some test cases withheld. Hugging Face

• Highlight: First public multi-institutional cine-MRI benchmark for real-time MRI-guided radiotherapy target tracking.

🏛️ FDA Clearances

• K261317 - Aidoc BriefCase Triage received 510(k) clearance for AI triage that prioritizes urgent imaging cases for review.

• K254120 - SubtleHD CT received 510(k) clearance for AI-based CT image enhancement aimed at improving image quality.

• K254013 - SubtleHD PET received 510(k) clearance for AI-based PET image enhancement to improve clarity and detail.

• K260378 - AZmed Rayvolve received 510(k) clearance for AI fracture detection support in musculoskeletal imaging.

• K252563 - Neurophet SCALE PET received 510(k) clearance for AI brain PET analysis to support neurologic imaging review.

• K261289 - Medtronic UNiD Spine Analyzer received 510(k) clearance for automated spine image analysis and surgical planning support.

• Explore last week's 16 radiology AI FDA approvals.

📄 Fresh Papers

• doi:10.3174/ajnr.A9140 - A prospective AJNR study in 76 MS patients found deep learning reconstructed 3D FLAIR detected all clinically relevant lesions.

• doi:10.1016/j.media.2026.104134 - An MRI-guided radiotherapy challenge used cine MRI from 585 patients at 6 institutions, with top trackers achieving Dice scores above 0.87.

• doi:10.1038/s41746-026-02798-w - A prospective NPJ Digital Medicine cohort used MRI and clinical data to predict rectal cancer pathologic complete response with AUC 0.790.

• doi:10.1186/s12880-026-02432-x - A prospective BMC Medical Imaging study found diastolic CT-FFR outperformed systolic CT-FFR, reaching AUC 0.885.

• Browse 174 new radiology AI studies from last week.

📰 Everything else in Radiology AI last week

• A simulation-trained MRI reconstruction method used 10% of diffusion data and could reduce brain MRI time from about 40 minutes to 8 minutes.

• Radiology Business reviews why Medicare AI payment remains split across inpatient, outpatient, and physician fee pathways.

• LLM support raised top-3 brain MRI diagnostic accuracy by 19.4 points for nonradiology residents and 14.7 points for radiology residents.

• A PET/CT model predicted Lu-177 PSMA absorbed dose in 9 patients, aiming to reduce the need for post-therapy dosimetry scans.

• Cloud-native AI architecture could centralize radiology software updates and speed deployment across imaging networks.

• Read 15 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.