Breast AI targets biopsy decisions across 27,048 patients

Good morning, there. BINDS achieved AUC 0.973 and could reduce benign biopsies by 32.4%.

I see this as a meaningful test of multimodal AI because it maps to how breast imaging escalates from ultrasound and mammography to MRI. The biopsy reduction claim matters only if local validation protects sensitivity and equity.

How much validation would you need before using AI to defer biopsy?

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

• Multimodal breast AI targets biopsy decisions

• Cardiac MRI model learns from reports, not manual labels

• AI speeds high risk mammography follow up

• NCCT AI lifts LVO reader sensitivity

• Plus: 2 newly released datasets, 6 FDA approved devices & 4 new papers.

🎗️ Multimodal breast AI targets biopsy decisions

RadAI Slice: A large multimodal breast AI study focused on reducing unnecessary biopsy while preserving diagnosis.

The details:

• Validated on 27048 participants across 8 centers

• Used ultrasound, mammography, MRI, and clinical data

• Reported AUC 0.973 for breast cancer risk assessment

• Could reduce benign biopsies by up to 32.4%

• Supported flexible modality combinations for different settings

Key takeaway: This feels clinically relevant because it targets biopsy decisions, not just image labels, but still needs prospective site level validation.

🫀 Cardiac MRI model learns from reports, not manual labels

RadAI Slice: This is notable because the training method could be more scalable than classic manual annotation.

The details:

• Model trained on more than 13000 cardiac MRI studies

• Used report text rather than manual image labels

• Performance reached up to 99% in selected tasks

• Testing across multiple institutions suggested generalizability

• Supported text-based search for cardiac MRI case retrieval

Key takeaway: For cardiac imaging teams, this suggests a practical path to specialty models built from existing report archives rather than costly annotation projects.

⚡ AI speeds high risk mammography follow up

RadAI Slice: This prospective deployment connects breast AI risk scoring with actual care delivery timelines.

The details:

• 4145 screening mammograms were scored in real time with Mirai

• Top 10% risk patients were flagged for expedited care

• 94% received immediate reads and 26 had same day diagnostics

• Timelines fell 99.1% for results and diagnostic evaluation

• Cancer detection reached 60 per 1000 in expedited patients

Key takeaway: The operational gain is the story here. AI moved from risk prediction to same day workflow triage in a safety net setting.

🧠 NCCT AI lifts LVO reader sensitivity

RadAI Slice: This stroke paper stands out because it tests AI on NCCT across countries and readers.

The details:

• Validated on 723 Korean and 240 US NCCT cases

• Standalone AUC was 0.963 in Korea and 0.899 in the US

• AI assistance raised pooled reader AUC from 0.718 to 0.852

• Sensitivity rose from 46.6% to 63.7%

• Specificity rose from 91.9% to 94.9%

Key takeaway: NCCT based LVO AI could support faster stroke triage, but the reader study frames it as assistance rather than replacement.

BIND (2026)

Modality: MRI, CT | Focus: Brain, CNS | Task: Classification, multimodal learning

• Size: 1,791,885 scans from 38,942 patients. Mostly MRI, with CT, PET, and SPECT.

• Annotations: Full-text imaging reports. LLM-extracted clinical categories and findings. MRI sequence labels. No manual segmentations.

• Institutions: Mass General Brigham, Stanford University

• Availability:

  restricted via BDSP

• Highlight: Large clinical neuroimaging dataset linked to EEG and PSG records, with LLM-derived report metadata.

Longitudinal-CT (2025)

Modality: CT | Focus: Whole-body; melanoma metastases | Task: Lesion segmentation; temporal tracking

• Size: 600 CT studies from 300 patients. Each patient has baseline and follow-up scans.

• Annotations: 7,182 expert manual tumor lesion segmentations. Labels include site, volume, and longitudinal correspondence/evolution.

• Institutions: University Hospital Tübingen, Fraunhofer MEVIS, et al.

• Availability:

  Public: FDAT DOI

• Highlight: Paired whole-body CT with exhaustive lesion masks and explicit lesion tracking across therapy timepoints.

🏛️ FDA Clearances

• K252954 - MammoSightAI received 510k clearance for lesion prioritization support in medical image review.

• K252558 - Philips Lumify received 510k clearance for portable diagnostic ultrasound imaging with Doppler support.

• K261132 - Healium Intelliscan LX192LC received 510k clearance for pulsed Doppler ultrasound imaging.

• K253837 - SPECTRALIS HRA plus OCT received 510k clearance for high resolution retinal imaging workflows.

• K260321 - HipGuide received 510k clearance as an image processing system for hip related radiology applications.

• K260716 - Neowise received 510k clearance as an image processing system for radiology visualization support.

📄 Fresh Papers

• doi:10.1148/ryai.250914 - DL single breath hold liver MRI synthesized full sequences from precontrast T1 and preserved HCC diagnosis performance.

• doi:10.1038/s41467-026-73170-5 - BIRD breast ultrasound AI was applied in 6817 screening patients across 107 hospitals with kappa 0.702.

• doi:10.1007/s00330-026-12610-8 - AI volumetric assessment of colorectal liver metastases added prognostic value beyond RECIST in liver only disease.

• doi:10.1093/bjr/tqag111 - An NHS multicenter evaluation of AI organ at risk contouring included 626 patients across 8 radiotherapy departments.

• Browse 191 new radiology AI studies from last week.

📰 Everything else in Radiology AI last week

• A mammography AI model predicted 10 year breast cancer risk with AUC 0.72 across KARMA and Olmsted cohorts.

• Seven commercial chest x ray AI tools showed sensitivity from 20.8% to 77.8% for lung cancer detection.

• Medical leaders proposed licensing style competency testing and supervised deployment for autonomous clinical AI.

• AMA and ACR backed caution around AI diagnosis claims, reinforcing human accountability in clinical care.

• A meta-review finds advanced models like DeepLabV3+ outperform U-Net for meningioma segmentation on MRI, but real-world deployment is limited by small tumor size.

• 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.