🩻 Prospective CXR AI cuts reading time by 18.3%

Good morning, there. A multicenter prospective chest x-ray study reduced interpretation time by 18.3%.

This week feels less about benchmark accuracy and more about deployment reality. One prospective CXR study shows measurable workflow gains, while ACR guidance gives imaging groups a practical framework for AI governance. At the same time, breast risk prediction and lung screening reviews show where AI may change clinical pathways, and where the evidence is still thin.

Is radiology AI finally moving from model performance to real workflow value?

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

• 🩻 Prospective CXR AI shows real workflow impact

• ACR sets a practical framework for imaging AI

• Mammography AI risk model beats density alone

• Lung screening AI still needs stronger evidence

• Plus: 6 FDA approved devices & 4 new papers.

🩻 Prospective CXR AI shows real workflow impact

RadAI Slice: This study stands out because it tested AI assistance in a prospective workflow, not just on a retrospective benchmark.

The details:

• Prospective multicenter deployment included 296 patients

• Interpretation time decreased by 18.3% compared with standard practice

• AI assistance improved report quality scores

• The system was designed to run on standard hardware, which may matter for lower resource settings

Key takeaway: For chest radiography, this is the kind of evidence radiology teams need: measurable workflow improvement, report quality signals, and a deployment setup that could be practical beyond highly specialized centers.

📘 ACR sets a practical framework for imaging AI

Image from: Radiology Business

RadAI Slice: The ACR guidance gives radiology groups a more concrete playbook for moving AI from purchase decision to supervised clinical operation.

The details:

• ACR Council approved the parameter in May 2024

• The framework was developed with SIIM

• It covers selection, monitoring, privacy, and ongoing performance

• The guidance connects with Assess AI registry work for quality monitoring

Key takeaway: The operational message is clear: imaging AI needs defined ownership, local validation, monitoring, and quality review. That makes adoption more standardized and defensible.

🎯 Mammography AI risk model beats density alone

RadAI Slice: This paper challenges the idea that breast density alone should drive supplemental screening decisions.

The details:

• Study included 123091 mammograms from 67019 women

• The DL model reached AUROC 0.71 versus 0.53 for density

• Dense breasts accounted for 41.4% of mammograms

• High risk exams had 2.1 false negatives per 1000

Key takeaway: The practical direction is toward image-derived risk stratification. For breast imaging, AI may help target supplemental MRI or ultrasound more precisely than density alone.

🫁 Lung screening AI still needs stronger evidence

RadAI Slice: This review is a useful reminder that product availability and clinical evidence are not the same thing.

The details:

• The review included 16 CE-marked lung nodule AI products

• Most products covered core tasks such as nodule detection and measurement

• Only 7% of 60 peer-reviewed studies were prospective

• No included studies reported patient outcomes or societal impact

Key takeaway: For lung screening programs, AI adoption should be paired with local monitoring, clear task definition, and realistic expectations about the current evidence base.

🏛️ FDA Clearances

• K260320 - Lunit INSIGHT MMG received FDA clearance to help detect suspicious cancer lesions on mammography for breast imaging workflows.

• K253163 - Ever Fortune AI received FDA clearance for a CT appendicitis triage system supporting faster notification in abdominal CT workflows.

• K260785 - Dentsply Sirona cleared an AI CBCT anatomy tool that automatically analyzes cone beam CT for dental imaging planning.

• K260009 - Broncus LungPoint VBN received clearance to support CT-based virtual bronchoscopic navigation for pulmonary procedures.

• K254237 - CliniComp cleared a PACS viewer for radiology image processing and image management workflows.

• K253077 - United Imaging cleared uOmnispace.MR, automated MR image processing software for radiology workflows.

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

📄 Fresh Papers

• doi:10.1002/mrm.70388 - Scanner-integrated machine learning generated whole-brain NODDI maps inline in under 10 seconds, addressing a key qMRI workflow barrier.

• doi:10.1007/s00330-026-12580-x - A review of 16 CE-marked lung screening AI products found task coverage gaps and only 7% prospective evidence across 60 studies.

• doi:10.1007/s00330-026-12595-4 - Prospective qCT analysis showed about 1.9% positional variability between supine and prone ILA scans, supporting consistent follow-up protocols.

• doi:10.1038/s41746-026-02712-4 - A review of 956 FDA radiology AI devices found limited closed-loop lifecycle management across adverse events, recalls, and software updates.

• Browse 154 new radiology AI studies from last week.

📰 Everything else in Radiology AI last week

• Radiology LLMs may help draft insurance appeal letters, but hallucinated references and variable accuracy still require physician review.

• SimonMed is selling AI add-ons for routine CT, MRI, and mammography, raising new questions about patient-pay imaging analytics.

• Healthcare AI leaders emphasized that governance, workflow integration, and infrastructure remain bigger adoption barriers than algorithms alone.

• A Stanford study found physicians used AI discharge summaries in 57% of cases, with 98% rated as low or extremely low harm risk.

• General multimodal LLMs reached 94.2% accuracy for scoliosis on x-rays but performed much less reliably on lumbar stenosis MRI.

• 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.