🩻 AI lifts CDR to 4.99 per 1000 DBT exams

Good morning, there. AI workflow raised generalist cancer detection from 3.76 to 4.99 per 1000 DBT exams.

I see this as a practical workforce story, not just an algorithm story. If generalists can safely detect more cancers, AI may help sites with limited breast subspecialty coverage. The recall increase still deserves close audit.

How would you monitor recall growth if deploying this workflow?

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

• 🩻 AI narrows the breast screening expertise gap

• Medicare AI prior auth pilot survives Senate vote

• AI prioritizes ED CT queues without new scanners

• CLEAR makes chest X ray AI more auditable

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

🩻 AI narrows the breast screening expertise gap

RadAI Slice: This prospective mammography study gives us a rare real world look at AI in deployed screening.

The details:

• Prospective study covered 577742 DBT screening exams at 109 US sites

• Generalists made up 60 of 95 radiologists and specialists made up 35

• Generalist CDR rose from 3.76 to 4.99 cancers per 1000 exams

• Recall rate rose from 9.06% to 10.40% while PPV also improved

• Specialist CDR stayed similar at 4.47 to 4.76 per 1000 exams

Key takeaway: For practices relying on generalists, this suggests AI plus safeguard review may narrow screening variation, though recall growth needs monitoring.

🏛️ Medicare AI prior auth pilot survives Senate vote

Image from: HealthExec

RadAI Slice: This policy fight matters because prior authorization can shape imaging access as much as technology.

The details:

• Senate effort to block CMS WISeR failed in a 46 to 50 vote

• Pilot uses AI for traditional Medicare claims in 6 states

• Vendors include Cohere Health, Humata Health and Innovaccer

• Opponents warned about care delays and denials for seniors

• Supporters framed WISeR as a waste and fraud reduction test

Key takeaway: Radiology groups should watch WISeR closely because AI authorization could affect imaging approvals, denials and appeal workload.

🚑 AI prioritizes ED CT queues without new scanners

RadAI Slice: This simulation stands out because it uses order time data to reshape CT wait times.

The details:

• Model trained on 313966 ED visits from 2020 to 2024

• Simulation tested 20795 CT studies from March to August 2024

• Median wait for actionable studies fell by 10.75 minutes

• 90th percentile wait fell by 43.36 minutes for actionable studies

• Actionable CTs within 1 hour rose from 48.33% to 57.30%

Key takeaway: Queue AI may give radiology a workflow lever for urgent findings, but sites need local simulation before changing dispatch rules.

🔎 CLEAR makes chest X ray AI more auditable

RadAI Slice: This foundation model is notable for making predictions traceable to radiology concepts.

The details:

• Trained on over 0.87 million image report pairs

• Training data came from 239391 patients

• Externally validated on 4 physician annotated datasets

• Predictions decompose into weighted radiologic observations

• Supports zero shot detection and confounder identification

Key takeaway: Auditable concept models could help radiologists inspect failure modes before deployment, especially in chest X ray triage and QA.

SynMo (2026-07-17)

Modality: CBCT | Focus: Head | Task: Motion simulation; motion compensation benchmarking

• Size: No image scans. 25 volunteers, 25 raw trajectories, 120 real 10-s windows, and 120 synthetic trajectories.

• Annotations: Ground-truth 6-DoF rigid head motion trajectories. JSON timestamps, translations, and Euler angles. No image labels.

• Institutions: Friedrich-Alexander-Universität Erlangen-Nürnberg; Siemens Healthineers AG

• Availability:

  Public: GitHub

• Highlight: Combines optically tracked real CBCT-style head motion with a retrainable VAE generator and pregenerated synthetic trajectories.

PhenSPINE (2026-07-22)

Modality: MRI | Focus: Spine, lumbar discs | Task: Multi-label classification, benchmarking

• Size: 16,813 DICOM images from 250 patients. 1,185 annotated IVD records across 4 MRI sequences.

• Annotations: IVD-level labels. Four binary pathology labels: herniation, bulging, spondylolisthesis, narrowing. Includes lumbar disc level metadata.

• Institutions: National Economics University, Phenikaa University, et al.

• Availability:

  Unspecified. Planned public release; no dataset link provided. Paper

• Highlight: Multi-sequence spine MRI benchmark with IVD positional metadata for pathology diagnosis.

OCT-Bench (2026-07-18)

Modality: OCT | Focus: Retina, choroid | Task: VQA, clinical reasoning

• Size: 4,137 OCT images; patient count not specified. 10,076 questions from 7 public datasets.

• Annotations: 10,076 expert-verified multiple-choice VQA labels. Uses standardized categories, boxes, masks, and clinical attributes from source datasets.

• Institutions: Shandong University; Joint SDU-NTU Centre for Artificial Intelligence Research; et al.

• Availability:

  Public: GitHub

• Highlight: Hierarchical OCT benchmark covering perception, cognition, and clinical reasoning across 20 tasks.

🏛️ FDA Clearances

• K261306 - Varian received 510k clearance for AI Contouring VA10A to automate structure delineation for radiation therapy planning.

• K253751 - Neurocareai received 510k clearance for CXRDetectAI, a chest X ray triage tool for lesion based prioritization.

• K253670 - Philips received 510k clearance for SmartCT R3.0, supporting advanced fluoroscopic imaging guidance during procedures.

• K260973 - Infervision received 510k clearance for InferOperate Suite, an imaging software tool for automated processing workflows.

• K262120 - Avara Software received 510k clearance for Avara Viewer, a radiology image viewing and analysis system.

• K253899 - Provect.AI received 510k clearance for 3D Anywhere, a radiology image processing system for 3D analysis workflows.

• Explore last week's 8 radiology AI FDA approvals.

📄 Fresh Papers

• doi:10.1093/esj/aakag085 - CATalyze AI shortened imaging to transfer request by 7.16 minutes and arrival to groin puncture by 31.28 minutes in stroke care.

• doi:10.1177/17474930261473812 - A multicenter synthetic CTA model generated CTA like images from NCCT in 3709 patients and reached 92.7% screening accuracy.

• doi:10.1016/j.ejrad.2026.113093 - An 18 institution CT pancreas model reached AUC 0.84 overall and AUC 0.97 for malignant lesions in external validation.

• doi:10.1016/j.jacr.2026.07.012 - A JACR thyroid ultrasound study found AI raised radiologist accuracy to 74.1% to 79.4% and cut FNA rates by 20.3% to 34.7%.

• Browse 249 new radiology AI studies from last week.

📰 Everything else in Radiology AI last week

• CARPL.ai raised $10 million to expand its marketplace of more than 300 radiology AI applications from over 100 partners.

• A JACR study found Qwen2.5 VL completed breast ultrasound report QC in 13 minutes for 50 reports versus 212.5 manually.

• Johns Hopkins and FDA collaborators introduced G AUDIT to detect spurious cues in medical AI training data before training.

• A CyberKnife dose prediction model cut computation from about 1 hour to 18 minutes in 251 retrospective cases.

• A healthcare AI trends report argued reimbursement should move toward outcomes as 86% of executives expect workforce change.

• 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.