CT AI screened 80,612 for esophageal cancer

Good morning, there. EAGLE validated noncontrast CT esophageal screening in 80,612 patients.

I see this as a rare screening study that tests AI on routine CT at real scale. If replicated, it could turn chest CT already in PACS into a referral signal for endoscopy, while keeping false positives visible to radiology teams.

Would you support CT based referral prompts for endoscopy in your workflow?

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

• Noncontrast CT AI enters esophageal screening

• FDA radiology AI passes 1094 devices

• LLM prompts improve report error detection

• Ultrasound AI targets BI RADS 4a biopsies

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

🦅 Noncontrast CT AI enters esophageal screening

RadAISlice: I would watch this because it tests AI on routine chest CT at screening scale.

The details:

• Validated across 80,612 patients from 12 centers in 3 countries

• External cohorts showed 98.5% specificity for opportunistic screening

• Cancer sensitivity was 90.0%, with 52.5% for precancerous lesions

• Prospective hospital validation reached 42.2% PPV in 17,446 patients

• Low dose screening reached 99.94% specificity in 10,959 people

Key takeaway: This could make routine chest CT a trigger for endoscopy referral, but radiology needs prospective workflow proof before adoption.

📊 FDA radiology AI passes 1094 devices

RadAISlice: This paper gives us a useful map of how radiology AI regulation is maturing.

The details:

• Study counted 1,094 FDA authorized radiology AI devices through 2025

• Annual authorizations rose 42% from 2023 to 2025

• Interpretive tools accounted for 21.5% of devices

• Cardiothoracic and neuro reached 70% of interpretive clearances

• PCCP authorization rose from 0.8% to 8.7% by 2025

Key takeaway: The field is scaling beyond detection tools, and change control plans may become central to safe radiology AI updates. We also maintain an up-to-date database of FDA-authorized radiology AI devices so you can follow the landscape as new products are added.

📝 LLM prompts improve report error detection

RadAISlice: This study stands out because it targets report QA rather than image diagnosis.

The details:

• Dataset included 1,170 clinician validated errors

• 1,200 reports covered radiography, ultrasound, CT and MRI

• RadCoT raised mean micro F1 from 0.77 to 0.85 across 7 models

• GPT 4o with RadCoT reached F1 0.93

• Interpretation error F1 rose from 0.57 to 0.75

Key takeaway: Structured prompts may make on premises report QA more realistic, especially for consistency errors radiologists often want surfaced.

🎗️ Ultrasound AI targets BI RADS 4a biopsies

RadAISlice: I like this one for its practical biopsy reduction endpoint in breast ultrasound.

The details:

• Prospective multicenter study enrolled 350 nodules from 331 patients

• Independent temporal test set included 90 nodules

• LightGBM model achieved AUC 0.95 with sensitivity 100%

• Specificity was 86.1%, with 32.4% of benign biopsies potentially avoided

• SHAP highlighted texture, penetrating vessel and BRCA status

Key takeaway: For breast imagers, the value is not replacing biopsy but creating a defensible second look for low risk 4a lesions.

CRCS-K Imaging Repository (2026)

Modality: CT, MRI, DSA | Focus: Brain, cerebrovascular | Task: Segmentation, detection

• Size: 225,159 imaging sequences from 20,792 patients. Acute ischemic stroke cases from 18 centers.

• Annotations: AI-derived features. Ischemic lesion volumes, perfusion metrics, WMH burden, microbleed counts, and LVO probabilities.

• Institutions: Seoul National University Bundang Hospital; JLK Inc.; et al.

• Availability:

  Request-only via AISCAN. Access requires steering committee approval and agreements.

• Highlight: Prospective sequence-level archive of all index-hospital stroke neuroimaging linked to clinical outcomes and AI quantification.

ABIDE II DTI Measurements (2026-09-10)

Modality: MRI (DTI) | Focus: Brain, white matter | Task: ASD classification, biomarker discovery

• Size: 151 DTI scans/subjects from ABIDE II. 150 valid after one failed control; 87 ASD and 64 controls.

• Annotations: ASD/control labels and image IDs. FA, AD, MD, and RD metrics averaged over 25 JHU white matter ROIs.

• Institutions: NYU Langone Medical Centre, San Diego State University, et al.

• Availability:

  Public: FAIR2 Data Portal

• Highlight: Ready-to-use, FAIR2-certified regional DTI features from ABIDE II with harmonization code.

UltraBench 2 (2026-09-23)

Modality: US | Focus: Multi-organ, prenatal | Task: Classification, segmentation

• Size: 12 datasets. 80,784 US frames/images from about 8,654 patients. 21 tasks.

• Annotations: Image-, patient-, and sequence-level labels. Lesion, organ, and anatomical segmentation masks.

• Institutions: UCLA; Université Laval et al.

• Availability:

  Restricted; code, data links, and leaderboard at GitHub. Some datasets are request-only.

• Highlight: Standardized ultrasound foundation model benchmark with fixed splits, preprocessing, and a public leaderboard.

MAMA-MIA (2025)

Modality: MRI | Focus: Breast | Task: Tumor segmentation; pCR prediction

• Size: 2,080 DCE-MRI exams from 2,080 patients. 1,506 training and 574 validation/test cases.

• Annotations: Voxel-wise primary tumor masks. Affected-breast bounding boxes and clinical metadata are included for training.

• Institutions: Universitat de Barcelona, Medical University of Gdansk, et al.

• Availability:

  Public training data; private validation/test. Website, CodaBench

• Highlight: Cross-continental benchmark with fairness scoring across age, menopausal status, and breast density.

UQMIA (Unspecified)

Modality: X-ray | Focus: Chest | Task: Uncertainty quantification; LLM evaluation

• Size: No new scans or patients. Includes 19 tutorial sessions and a 100-item MCQ bank.

• Annotations: MCQ answer labels. Uses an external pediatric chest radiograph dataset for examples.

• Institutions: Mashhad University of Medical Sciences, Mayo Clinic, et al.

• Availability:

  Public: website and GitHub

• Highlight: Open hands-on UQ tutorial for medical imaging with executable Kaggle notebooks and LLM-based content assessment.

SonoCorpus (2026-09-16)

Modality: US | Focus: Multi-organ; fetal/cardiac | Task: Segmentation; clinical measurement

• Size: 456,963 images/frames/slices from 53 public datasets. Patient count not reported.

• Annotations: 1,626,085 expert segmentation masks. Organ, lesion, nerve, muscle, vessel, and fetal structures.

• Institutions: MBZUAI; ADIA Lab; et al.

• Availability:

  Public manifest. Images and masks are accessed from original sources. Zenodo

• Highlight: Largest open ultrasound segmentation corpus to date. Spans 24 clinical applications, 17 countries, and 2D/video/3D US.

🏛️ FDA Clearances

• K252486 - Gleamer cleared ChestView US 1.1 for AI analysis of chest ultrasound findings.

• K261795 - Quicktome Software Suite clears automated brain image processing and neuroanatomic analysis support.

• K254072 - Curvebeam AI cleared BoneVue for imaging based bone density measurement and MSK assessment.

• K260133 - Odin Medical cleared CADDIE to support gastrointestinal lesion detection from medical images.

• K251417 - Argus Cognitive cleared ReVISION 2 for AI aided analysis of radiological images.

• K260879 - INVIA cleared 4DM software for automated radiological image processing and clinician review support.

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

📄 Fresh Papers

• doi:10.1016/j.jcct.2026.09.005 - In 1,193 ED chest pain patients, CCTA AI had ICA sensitivity 94.6% and NPV 89.2% for obstructive CAD.

• doi:10.1007/s10140-026-02546-3 - Extremity fracture AI showed 92.7% sensitivity and 94.4% specificity in 2,508 radiograph patients.

• doi:10.1093/noajnl/vdag236 - FDA cleared VBrain tracked 322 brain metastases in CROWN and improved PFS prediction by c index 0.07.

• doi:10.1148/rycan.250609 - In 20,187 mammograms, detection and risk AI models performed best within their intended use settings.

• Browse 289 new radiology AI studies from last week.

📰 Everything else in Radiology AI last week

• ORACLE used CT and clinical data in 2,746 patients and reached external AUC 0.81 with 97.2% NPV.

• CIPHER used baseline CT from 2,500 lung cancer patients to predict pneumonitis with external AUC about 0.83.

• BCRF launched a $100M fund and made Clairity, an FDA De Novo mammography risk tool, its first investment.

• FDA clearance lets Abdo Triage flag up to 17 abdomen pelvis CT findings for workflow prioritization.

• AI enhanced handheld carotid ultrasound detected 94.8% of plaques versus 87.6% with standard images.

• Read 14 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.