93% sensitivity in FDA-cleared AI for CT lung cancer screening

RadAISlice Newsletter

Weekly Updates in Radiology AI

Good morning, there. The FDA cleared Eyonis LCS with 93.3% sensitivity and 92.4% specificity for CT lung screening.

I see this as a strong step toward credible, scalable AI in lung cancer screening. Robust metrics and direct FDA clearance as an end-to-end device for LDCT workflows should help practices address reader variability and rising demand. The system’s external validation should also support broader adoption across diverse sites and patient groups.

How would you integrate FDA-cleared AI into your CT screening workflow?

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

• FDA clears AI tool for CT lung cancer screening with high accuracy

• Mass General unveils scalable neuroradiology AI foundation model

• Deep learning predicts GIST mutations and outcomes across 7 countries

• AI-driven prenatal ultrasound detects placenta accreta with 100% sensitivity

• Plus: 1 newly released dataset, 4 FDA approved devices & 4 new papers.

Will you attend ECR 2026 in Vienna?

🫁 FDA clears AI tool for CT lung cancer screening with high accuracy

Image from: Radiology Business

RadAISlice: FDA clearance of Eyonis LCS establishes a high-performing AI solution for LDCT lung screening.

The details:

• 93.3% sensitivity in testing for lung cancer detection

• 92.4% specificity and 99.9% NPV reported

• First FDA-cleared comprehensive AI tool for lung screening workflow

• Designed to assist with reader variability and screening access

Key takeaway: Widespread validated AI in lung cancer screening could help standardize interpretation, handle higher volumes, and improve patient outcomes.

🧠 Mass General unveils scalable neuroradiology AI foundation model

Image from: Radiology Business

RadAISlice: BrainIAC marks a shift from narrow to versatile neuroimaging AI for MRI applications.

The details:

• Foundation model adapts to diverse neuroradiology tasks

• Self-supervised learning across large unlabeled MRI sets

• Designed for generalization and external population robustness

• Outcomes published in Nature Neuroscience

Key takeaway: Moving from task-specific to foundation models could improve generalizability and personalization in neuroimaging AI, prompting renewed focus on robust validation across populations.

🧬 Deep learning predicts GIST mutations and outcomes across 7 countries

RadAISlice: This multinational study shows how DL on whole-slide images predicts key clinical variables for GIST.

The details:

• 7,238 molecular, 2,638 follow-up cases, 21 centers, 7 countries

• AUCs: 0.87 (KIT), 0.96 (PDGFRA), HR for RFS prediction: 8.44

• Performance was comparable to pathology risk scores

• Models explainable via visual heatmaps and tiles

Key takeaway: International validation of DL pathology in GIST demonstrates it can add clinically relevant mutation and prognostic predictions, supporting future multimodal AI strategies.

🍼 AI-driven prenatal ultrasound detects placenta accreta with 100% sensitivity

RadAISlice: AI-powered ultrasound detected all PAS cases in a high-risk prenatal cohort without missing any true cases.

The details:

• Retrospective study of 113 pregnancies at Texas Children’s (2018–2025)

• Zero false negatives and 2 false positives for PAS detection

• Mean diagnosis: 30.9 weeks gestation

• Current screening misses ∼50% of cases

Key takeaway: Specialized ultrasound AI for placenta accreta could support earlier intervention and reduce maternal morbidity in obstetric imaging.

CT-RATE (2024-02-08 (arXiv v5, public release, referenced as 2025 in dataset repository))

Modality: CT | Focus: Chest | Task: Report generation, multi-abnormality detection

• Size: 25,692 3D scans from 21,304 patients, >14.3M slices

• Annotations: Paired free-text radiology reports (impression & findings); global labels for 18 abnormalities (automatic + some manual annotation); anatomical segmentation for organs

• Institutions: Istanbul Medipol Univ., Univ. of Zurich et al.

• Availability:

  Public (Hugging Face link)

• Highlight: Largest open 3D chest CT dataset with paired radiology reports; supports foundation model training and zero-shot tasks

🏛️ FDA Clearances

• K251769 - RevealAI-Lung (Precision Medical Ventures) improved malignancy discrimination (AUC +0.18, p<0.0001) for incidental lung nodules in a multi-reader, multi-case study.

• K252496 - Neurophet AQUA AD Plus achieved Dice 0.83–0.90 for brain MRI and PET structure quantification, with SUVR analysis matching FDA-cleared reference systems (>0.99 ICC).

• K251195 - Aidoc's BriefCase-Triage for CT brain aneurysm triage (≥3mm) demonstrated 87.8% sensitivity and 91.6% specificity in multicenter, blinded validation.

• K253499 - Eyas Medical's Ascent3T Neonatal MRI system for infants <9kg completed verification/validation testing and supports detailed anatomical imaging at the point of care.

• Explore last week's 9 radiology AI FDA approvals.

📄 Fresh Papers

• doi:10.64898/2026.02.07.26345811 - Stanford's TRIAD-HFpEF framework combines MRI, ECG, and biomarkers in UK Biobank, finding 90+ loci and actionable targets for HFpEF through ML phenotyping.

• doi:10.1016/j.radi.2026.103348 - A systematic bibliometric review traces two decades of deep learning growth and clinical validation in AI-augmented TB imaging, emphasizing large-scale chest X-ray use.

• doi:10.1038/s41467-026-68414-3 - A multimodal AI (n=2200+) for liver steatosis and fibrosis demonstrated AUCs 0.90–0.93 (steatosis) and 0.82–0.89 (fibrosis), enabling early stratification during opportunistic CT scans.

• doi:10.64898/2026.02.04.26345495 - The SCOPE model improved early small pancreatic lesion detection on CT, raising sensitivity from 0.60 to 0.73 at 95% specificity versus segmentation baselines, in a 3-cohort study.

• Browse 135 new radiology AI studies from last week.

📰 Everything else in Radiology AI last week

• Columbia University experts warn unchecked AI deployment in medical imaging could lead to ineffective 'zombie algorithms' and workflow risks.

• A large study shows popular LLMs often repeat fabricated or unsafe medical claims found in clinical and social media notes, underscoring validation gaps.

• Despite >1,000 FDA-cleared radiology AI tools, BLS projects 5% radiologist job growth through 2034, supporting an augmentation not replacement narrative.

• A 10-center, 204-patient study found Canon's deep learning super-resolution improved CCTA stenosis AUC from 0.90 to 0.97 over hybrid iterative methods.

• 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.