AI reads pancreatitis on the first CT

Good morning, there. APEX NET reached AUC 0.981 for acute pancreatitis diagnosis on noncontrast CT.

I like this one because it tackles the first CT many emergency patients receive rather than waiting for contrast. Earlier severity stratification could support triage, consults, and follow up imaging decisions.

Would you use NCCT AI to guide early pancreatitis triage?

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

• AI reads early pancreatitis on noncontrast CT

• Stroke AI adoption tracks resources

• BoneCoT scales CT bone metastasis assessment

• Structured AI reports speed CCTA reading

• Plus: 5 newly released datasets, 6 FDA approved devices & 4 new papers.

🚨 AI reads early pancreatitis on noncontrast CT

RadAI Slice: This study caught my eye because it targets the first CT many pancreatitis patients actually get.

The details:

• 3383 patients across 5 centers, with retrospective and prospective cohorts

• AP diagnosis AUC reached 0.949 to 0.981 across validation and test cohorts

• Severity prediction macro AUC was 0.873 in validation and 0.872 internally

• Reader study found performance comparable to senior radiologists

Key takeaway: For emergency abdominal CT, this could move AI from detection toward early severity triage before contrast CT is available.

💸 Stroke AI adoption tracks resources

Image from: Radiology Business

RadAI Slice: This Medicare analysis shows why reimbursement alone may not equal equitable AI access.

The details:

• 5 percent Medicare claims sample covered 2020 to 2023 stroke episodes

• AI use peaked at 21 percent in 2022, then fell after NTAP ended

• Use was 6 times higher in 2022 and 2 times higher in the Stroke Belt

• Deprived area hospitals were significantly less likely to use the tool

Key takeaway: Adoption work needs integration, staffing, and trust support, not just payment codes, if stroke AI is to reach smaller hospitals.

🦴 BoneCoT scales CT bone metastasis assessment

RadAI Slice: This foundation model stands out for its scale and multisite oncology scope.

The details:

• Pretrained on 29.3 million CT images from 30,267 patients

• Covered 12 skeletal sites and 26 clinically relevant tasks

• Tested across 10 hospitals and multicenter cohorts

• Reported 20 percent AUC gain versus state of the art methods

• Primary versus metastatic lesion AUC improved by 40 percent

Key takeaway: For oncologic CT, the study points toward broader lesion context, but prospective workflow validation still matters.

🫀 Structured AI reports speed CCTA reading

RadAI Slice: This workflow study feels practical because it measures reporting time and agreement.

The details:

• 80 CCTAs were read by 5 readers across 2 clinical sites

• Structured tool cut reporting time 40.2 percent, from 10.0 to 6.0 minutes

• Very confident reads rose from 29 percent to 48 percent

• Report agreement increased from 45.3 percent to 94.6 percent

• Cognitive load fell from 6.1 to 3.8 on a 9 point scale

Key takeaway: CCTA AI may have near term value when it enters the report, not just the image, but larger real workflow studies are needed.

EUCAIM HUVR LABC PET-CT (2026; exact dataset release date unspecified)

Modality: PET-CT | Focus: Breast, axillary lymph nodes | Task: NST response prediction, federated learning

• Size: About 200 PET-CT studies from 100 patients; baseline and pre-surgical scans.

• Annotations: Patient-level clinical-demographic variables and NST response data. No segmentations reported.

• Institutions: Virgen del Rocío University Hospital, IBiS/CSIC/US, et al.

• Availability:

  Restricted via EUCAIM federated infrastructure; data remain locally stored. EUCAIM

• Highlight: FAIR multimodal PET-CT and clinical dataset mapped to the EUCAIM Common Data Model for federated AI.

CRL-2025 (2025)

Modality: MRI (T2w, DTI) | Focus: Fetal brain | Task: Segmentation, parcellation

• Size: 193 reconstructed T2w scans from 159 fetuses. T2 atlas spans 21–37 GA; DTI atlas spans 23–38 GA.

• Annotations: 31 tissue labels and 126 regional parcellations. Includes transient WM compartments and DTI labels.

• Institutions: University of California Irvine, Boston Children’s Hospital et al.

• Availability:

  Public: Harvard Dataverse

• Highlight: 4D fetal brain atlas with detailed tissue labels, transient WM compartments, and a released FetalSEG model.

BEAMSTER (1 July 2026)

Modality: MRI | Focus: Brain, brain metastases | Task: Detection, segmentation

• Size: 140 contrast-enhanced T1w MRI scans from 140 patients. 260 metastatic lesions.

• Annotations: Expert binary NIfTI masks of brain metastases. Derived from GTV/PTV RT contours.

• Institutions: University Hospital & Faculty of Medicine Ostrava; Brno University of Technology

• Availability:

  Public: Figshare. CC-BY 4.0.

• Highlight: Enriched with small brain metastases from stereotactic radiotherapy planning.

HealthAgentBench (2026-06-30)

Modality: X-ray, CT, WSI, EHR | Focus: Chest; EHR | Task: Agent evaluation; diagnosis

• Size: 54 tasks. Imaging subset: 10 CXR cases, 10 CT volumes, 10 WSI slides. EHRSHOT source has 6,739 patients.

• Annotations: Hidden gold labels and verifiers. Includes expert-reviewed imaging labels, tumor masks, report corrections, trial qrels, EHR errors, and AUROC baselines.

• Institutions: Microsoft Research; PhysioNet et al.

• Availability:

  Public benchmark code; data mixed public and gated. GitHub

• Highlight: A unified agentic healthcare benchmark with executable Docker environments and raw multimodal clinical data.

CORTEX (2026-06-27)

Modality: CT | Focus: Chest, lungs | Task: VQA, report generation

• Size: 3,039 CT examinations from 1,304 patients; 76,177 validated reasoning traces.

• Annotations: Validated four-stage reasoning traces for VQA and report generation. Includes clinical context and final answers; no segmentations.

• Institutions: MBZUAI, Hasso Plattner Institute, et al.

• Availability:

  Unspecified; planned public release upon acceptance. Code

• Highlight: Structured, stage-level verified reasoning for 3D chest CT, with patient history reattached.

🏛️ FDA Clearances

• K260729 - GuideAI Health received 510k clearance for AI software that triages suspected vascular occlusions.

• K253690 - Siemens Healthineers received 510k clearance for LungMaps, a CT lung analysis and segmentation tool.

• K253825 - DeepHealth received 510k clearance for Saige Dx, AI software for suspicious lesion detection.

• K260300 - Anumana received 510k clearance for WatchMate Software, an AI radiology image processing tool.

• K261713 - FUJIFILM received 510k clearance for Synapse PACS 7.6.0, updating a major imaging workflow platform.

• K253264 - Siemens received 510k clearance for myAblation Guide, software for image based ablation planning.

• Explore last week's 13 radiology AI FDA approvals.

📄 Fresh Papers

• doi:10.3174/ajnr.A9504 - A MRMC AJNR study found automated ASPECTS support improved balanced accuracy by 5.7 points on 100 NCCT stroke cases.

• doi:10.1097/RLI.0000000000001300 - A prospective MS MRI study found DL accelerated 3D T2 imaging was interchangeable while cutting scan time nearly 50 percent.

• doi:10.1148/radiol.251821 - SCOT HEART analysis linked AI derived CCTA body composition, especially muscle attenuation, with 10 year MI and mortality.

• doi:10.1161/CIRCIMAGING.125.019726 - A 10 center study of 29,339 MPI CT attenuation scans found AI aortic size indices predicted mortality beyond perfusion data.

• Browse 331 new radiology AI studies from last week.

📰 Everything else in Radiology AI last week

• An Irish imaging center survey found 96 percent of patients want disclosure when AI is used in imaging reports.

• A breast imaging review reported 29 percent more cancers with AI assisted mammography and about 40 percent less reading time.

• vRad presented work on generative AI for full chest X ray report drafting and quality assurance.

• Read 8 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.