CMR AI raises mortality C index by 0.033

Good morning, there. AI derived CMR ejection fraction improved mortality C index by 0.033 in 27,382 patients.

For me, this makes AI quantification feel less like a speed tool and more like clinical risk information. The effect size is modest, but the cohort size and outcome link make it relevant to daily CMR workflow.

How much prognostic value should influence adoption of CMR quantification AI?

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

• CMR AI links EF to survival at scale

• FDA AI authorizations keep radiology in front

• Lung screening AI combines risk and nodule models

• AI guided echo reaches rural community scans

• Plus: 2 newly released datasets, 5 FDA approved devices & 4 new papers.

🫀 CMR AI links EF to survival at scale

RadAISlice: This study makes AI EF feel less like automation and more like risk information.

The details:

• 27,382 CMR patients from UK and US cohorts

• 3458 deaths occurred over median 5.6 years

• AI EF C index was 0.669 vs 0.637 for clinician EF

• After age and sex adjustment, C index was 0.781 vs 0.777

• AI added to clinician EF improved model fit with p under 0.0001

Key takeaway: The result supports first line AI EF measurement with oversight, especially where CMR volume and consistency pressure are high.

📊 FDA AI authorizations keep radiology in front

RadAISlice: This regulatory snapshot helps separate market scale from actual clinical adoption.

The details:

• Annual authorizations rose from 18 in 2016 to 335 in 2025

• 977 of 1,614 entries were authorized from 2023 onward

• Radiology accounted for 1,230 entries, or 76.2%

• 510(k) clearance accounted for 1,553 entries, or 96.2%

• Imaging systems made up 399 entries, or 24.7%

Key takeaway: Radiology remains the FDA AI proving ground, but embedded scanner AI and post market surveillance now need closer scrutiny.

🫁 Lung screening AI combines risk and nodule models

RadAISlice: This study stands out because it tests AI across two screening cohorts, not one curated dataset.

The details:

• 5,110 baseline LDCT exams included 65 cancers within 1 year

• Sybil reached AUC 0.93 in both NLST and P IELCAP

• MONAI AUC was 0.90 in NLST and 0.79 in P IELCAP

• Composite model sensitivity was 96.9%, missing 2 of 65 cancers

• Positive predictive value was 2.7% at the selected operating point

Key takeaway: High sensitivity may be useful for safety net triage, but low PPV means workflow design matters as much as model choice.

🩺 AI guided echo reaches rural community scans

RadAISlice: This access story is compelling because it tests acquisition support outside specialist imaging sites.

The details:

• 4,597 adults were scanned in rural Alabama, Kentucky, Louisiana and Mississippi

• 14 non sonographers trained with a 2 hour AI guided curriculum

• 98% of scans included all 6 standard echo views

• Core lab experts rated more than 99% of scans adequate

• 133 critical results were detected, with 99% returned within 3 days

Key takeaway: AI guided acquisition could extend imaging reach, while preserving radiology and cardiology oversight through remote review.

MedHEB (2026-11-02)

Modality: CT, MRI, X-ray, text | Focus: multi-organ, chest/thorax | Task: retrieval, embedding evaluation

• Size: 75,233 queries across 107 tasks from 54 source datasets. Pooled scan and patient counts are not reported.

• Annotations: Source labels, paired reports, VQA answers, cross-modal pairs, and region crops/bounding boxes.

• Institutions: The University of Sydney, Shandong Medical and Pharmaceutical University, et al.

• Availability:

  Unspecified. Authors state release of task indices, preprocessing scripts, evaluation code, and access manifests; no dataset URL is provided.

• Highlight: Unified benchmark for 2D, 3D, and text medical retrieval, including modality-bridging and region-level tasks.

MedNet (2026-11-01)

Modality: X-ray, MRI, CT, etc. | Focus: multi-organ; human and veterinary | Task: text-to-image generation; diagnostic classification

• Size: 241,598 images. Patient count not specified. 193,163 train and 48,435 test images.

• Annotations: 160 disease class labels across 61 categories. Includes text descriptions per class. No segmentations reported.

• Institutions: Yeshiva University, Boston University, et al.

• Availability:

  Public via GitHub

• Highlight: Large multi-modal dataset spanning human and animal diseases for broad medical image generation.

🏛️ FDA Clearances

• K262062 - Avicenna.Ai received 510(k) clearance for CINA Aneurysm, a CT AI tool for brain aneurysm detection.

• K262147 - Avicenna.Ai received 510(k) clearance for CINA Aneurysm Plus, a CT AI tool for brain aneurysm detection.

• K262080 - Cascination received 510(k) clearance for OTOPLAN, software for image based otologic and neurotologic planning.

• K261531 - Perception Vision Medical Technology received 510(k) clearance for PV iCurve Pro image processing software.

• K261679 - Deep Breathe received 510(k) clearance for BlineSlide, AI software intended to help detect suspicious lesions.

📄 Fresh Papers

• doi:10.1007/s00330-026-13028-y - AI fracture detection found 21 of 26 compensated missed fractures and could have mitigated 93.8% of payouts.

• doi:10.1016/j.ajogmf.2026.102192 - A five hospital randomized trial found AI feedback improved fetal Doppler assessment accuracy from 62.3% to 75.0%.

• doi:10.1186/s13244-026-02428-9 - A multicenter MRI radiomics nomogram reached external AUC 0.829 for FIGO 2023 endometrial cancer staging.

• doi:10.1016/j.jcmg.2026.08.019 - A multicenter CTA deep learning model quantified aortic valve fibrosis with expert ICC 0.913 and histology ICC 0.905.

• Browse 287 new radiology AI studies from last week.

📰 Everything else in Radiology AI last week

• An open source pediatric imaging platform handled 500 clinical AI tasks with 99.2% processing success.

• An AI enabled Medicare prior authorization pilot was linked to 15% to 20% fewer targeted services in six states.

• An fMRI machine learning tool trained on SuperAgers and Alzheimer patients predicted decline and amyloid accumulation.

• A large neural network reproducibility study found one bit changes altered outcomes in nearly 48% of cases.

• Read 11 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.