Mammography AI tested on 707922 exams

Good morning, there. A Kaiser study compared mammography AI on 707922 screening exams.

I see this as a rare scale test of how breast AI behaves before adoption decisions. The spread across algorithms makes local validation and threshold choice feel less optional. It also reminds us that workflow intent matters as much as AUC.

How would you validate mammography AI before using it in screening?

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

• Mammography AI gets a 707922 exam reality check

• AI workflow gains hinge on routing speed

• Open source PET AI scales lesion segmentation

• CLAIM turns AI reporting into a 44 item playbook

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

🎗️ Mammography AI gets a 707922 exam reality check

RadAISlice: This study stands out because it tests commercial breast AI at screening scale.

The details:

• 707922 exams with complete data and 4436 cancers within 12 months

• Radiologists recalled 6.8 percent with 67.2 percent sensitivity

• Best AI AUROC was 0.849, ahead of 0.818, 0.817 and 0.815

• No AI matched radiologist sensitivity at AI positive rates of 10 percent or less

Key takeaway: This changes breast AI procurement from vendor claims to local operating point validation before workflow deployment.

⚙️ AI workflow gains hinge on routing speed

RadAISlice: The result feels practical because it measures what radiologists actually experience.

The details:

• 10 AI tools from 7 vendors ran across 20 outpatient imaging centers

• Technical cohort included 96874 exams over 4.5 years

• Median latency was 2.06 minutes, with 72 percent from data routing

• AI availability cut TAT by 26 percent for trauma x ray and 18 percent for knee MRI

Key takeaway: For deployed AI, PACS routing and result timing may determine clinical value more than algorithm runtime.

🧬 Open source PET AI scales lesion segmentation

RadAISlice: This nuclear medicine study is notable for scale, openness and modality specific design.

The details:

• Training used 5209 FDG PET CT scans across 19 disease types

• PSMA targeted PET CT training included 2046 scans

• Organ context raised median Dice from 0.62 to 0.71 for FDG

• PSMA median Dice reached 0.85 and beat other open source tools

Key takeaway: Open PET segmentation tools could support quantitative oncology workflows if sites verify scanner and disease mix.

📋 CLAIM turns AI reporting into a 44 item playbook

RadAISlice: This guidance matters because weak reporting still limits trust in imaging AI evidence.

The details:

• CLAIM 2024 includes 44 reporting items across the AI study lifecycle

• The update followed Delphi consensus with 72 imaging and AI experts

• The paper explains common misreadings and gives literature examples

• An online tool supports authors, reviewers and editors

Key takeaway: Better reporting will not validate AI by itself, but it raises the floor for reproducible radiology AI evidence.

GE 11L-D Multi-anatomical Ultrafast US Dataset (15 September 2026)

Modality: US | Focus: Multi-anatomical; phantom | Task: Beamforming; image reconstruction

• Size: 20,431 acquisitions: 20,121 in vivo from 16 volunteers, plus 310 CIRS phantom acquisitions.

• Annotations: Raw RF data with metadata labels. Body region, volunteer ID, and phantom setting; no lesion labels or segmentations.

• Institutions: EPFL

• Availability:

  Public: Zenodo

• Highlight: Large open raw RF ultrafast US dataset with 105 plane waves per acquisition across six body regions.

RSNA-ICA (2025)

Modality: CTA, MRA, MRI | Focus: Brain, intracranial vasculature | Task: Aneurysm detection/localization, vessel segmentation

• Size: 7202 series from 4278 adult patients. Public release includes 5041 series.

• Annotations: Expert labels for aneurysm presence and 13 vascular locations. Crosshair localizers. 3D arterial segmentations for 178 series.

• Institutions: RSNA, ASNR, et al.

• Availability:

  Public for noncommercial use: MIRA. Challenge training set also on Kaggle.

• Highlight: Largest diverse public intracranial aneurysm dataset, spanning CTA, MRA, and routine MRI from 21 institutions across five continents.

FOCUS (2026-09-27)

Modality: CFP | Focus: Retina, optic disc | Task: Disease classification, model benchmarking

• Size: 10 public fundus datasets. Total scans and patients are not specified.

• Annotations: Harmonized binary labels for any DR, referable DR, and GON. Age, sex, and image-quality metadata where available.

• Institutions: CentraleSupélec, University of Michigan, et al.

• Availability:

  Public benchmark; source datasets may have mixed licenses. Link unspecified.

• Highlight: Cross-dataset benchmark for retinal foundation models under dataset shift, with calibration, fairness, and quality robustness analyses.

🏛️ FDA Clearances

• K261517 - Qure.ai qCT Lung cleared for CT based lung lesion detection support in suspected lung cancer workflows.

• K260906 - A2z Radiology a2z Abdo Triage cleared for AI aided urgent abdominal CT triage.

• K261855 - Elucid PlaqueIQ cleared for CT based vascular plaque image analysis support.

• K261677 - AIRA Version 1 cleared for AI renal assessment on CT imaging.

• K261623 - GE EchoPAC Software Only and EchoPAC Plug in v211 cleared for ultrasound image analysis workflows.

• K261861 - SPINEsim v1.0 cleared for spine image processing and planning support.

• Explore last week's 7 radiology AI FDA approvals.

📄 Fresh Papers

• doi:10.1007/s00330-026-12888-8 - A multicenter prostate MRI study found AI patient level sensitivity of 97.6 percent versus 92.6 percent for PI RADS.

• doi:10.64898/2026.07.16.26356882 - Prospective AI guided DVT ultrasound reached 92.86 percent triage sensitivity across 594 subjects.

• doi:10.1002/jmri.70575 - A liver MRI prescription model validated on 11012 patients achieved median axial IoU of 0.963.

• doi:10.1016/j.jacadv.2026.103277 - Automated aortic valve calcium scoring on nongated chest CT reached r 0.99 and 0.92 sensitivity for AS burden.

• Browse 316 new radiology AI studies from last week.

📰 Everything else in Radiology AI last week

• Academic Radiology analysis found 1094 FDA authorized radiology AI devices through December 2025.

• Stockholm will use Lunit AI as an independent second reader for 200000 to 250000 annual mammograms.

• Harrison.ai cut 15 percent of staff while reporting 28 million dollars in ARR and US teleradiology expansion.

• GE HealthCare named Rodolphe Katra global chief AI officer as it expands its imaging AI strategy.

• Read 10 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.