CT AI flagged occult pancreatic cancer 475 days early

Good morning, there. REDMOD detected occult pancreatic cancer on CT a median 475 days before diagnosis.

I see this as a serious test of opportunistic CT AI, not a screening claim. The external validation and specificity data matter because false positives could overwhelm radiology and GI pathways. It also sharpens the question of where opportunistic AI belongs in abdominal CT workflow.

Would your practice act on an AI flag for pancreatic cancer 16 months early?

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

• CT AI spots occult pancreatic cancer early

• ACR builds real world AI monitoring

• Chest CT AI flags acute heart failure

• AI helped mammography readers cut time 32%

• Plus: 2 newly released datasets, 3 FDA approved devices & 4 new papers.

🧬 CT AI spots occult pancreatic cancer early

RadAISlice: A multi institution CT radiomics model found visually occult pancreatic cancer more than a year earlier.

The details:

• Trained on 969 CTs and tested on 493 independent CTs

• Detected occult PDAC with AUC 0.82 and 73.0% sensitivity

• Radiologist sensitivity was 38.9% overall and 23.0% beyond 24 months

• Median lead time was 475 days before clinical diagnosis

• Specificity reached 81.3% and 87.5% in external cohorts

• Prospective high risk workflow validation is still needed

Key takeaway: This could shift CT review toward pancreatic risk triage, but adoption should wait for prospective validation and clear follow up pathways.

📊 ACR builds real world AI monitoring

RadAISlice: This registry is a practical answer to one of radiology AI’s hardest problems: performance drift after deployment.

The details:

• Assess AI runs within the ACR NRDR using ACR Connect

• Sites submit AI output, report text and DICOM metadata

• LLM prompts extract surrogate labels from reports

• Use cases include ICH, PE, pneumothorax, LVO and bone age

• Dashboards track concordance, benchmarks and discordance

Key takeaway: For radiology groups, this moves AI governance from one time validation to ongoing surveillance with local case review.

🫀 Chest CT AI flags acute heart failure

RadAISlice: This prospective CT study stands out because dyspnea workups already strain ED radiology.

The details:

• 234 dyspneic ED patients had low dose noncontrast chest CT

• 61 patients had radiologist reported acute heart failure

• AI reached AUROC 0.95 with 89% sensitivity and specificity

• Rule out setting reached 97% sensitivity and 74% specificity

• Performance was comparable to radiologists and cardiologists

Key takeaway: For ED chest CT, AI could standardize pulmonary congestion calls and support triage when specialist reads are delayed.

🧠 AI helped mammography readers cut time 32%

RadAISlice: I like that this study focused on workflow and specificity rather than only headline accuracy.

The details:

• Nine radiologists read 302 mammograms with and without AI

• Specificity rose from 77.0% to 88.4%

• AUC improved from 0.799 to 0.851

• Read time fell from 121.5 to 83.2 seconds per case

Key takeaway: This is the kind of operational evidence that can matter in screening programs, especially where staffing pressure makes every minute count.

FibreSimulator (2024 (v1.0.0))

Modality: CT | Focus: FRP composites; fibres | Task: Reconstruction; segmentation

• Size: Synthetic on-demand volumes. Examples include 256³ voxels with 300–500 fibres and 800³ voxels with 6000 fibres. 0 patients.

• Annotations: Voxel-level ground truth for fibres, resin and air. Includes defects such as holes, notches and voids.

• Institutions: Universiteit Leiden; University of Manchester

• Availability:

  Public: Zenodo

• Highlight: Open-source generator for labelled 3D fibre phantoms with CT simulation via ASTRA.

BraTS-PEDs (2026)

Modality: MRI | Focus: Brain, pediatric high-grade glioma | Task: Tumor segmentation, benchmarking

• Size: 457 mpMRI studies from 457 pediatric patients. Training 257, validation 91, testing 109.

• Annotations: Expert-refined 3D masks for ET, NET, cystic component, and edema. RAPNO-aligned.

• Institutions: Children’s Brain Tumor Network, International DMG/DIPG Registry, et al.

• Availability:

  Public via TCIA: BraTS-PEDs

• Highlight: Largest public multi-institutional mpMRI benchmark for pediatric brain tumor segmentation.

🏛️ FDA Clearances

• K253502 - GE received 510(k) clearance for Critical Care Suite software that detects enteric feeding tube position on radiographs.

• K252237 - Edgecare received 510(k) clearance for EdgeFlow UW20, an ultrasonic imaging system for diagnostic visualization.

• K252587 - Shenzhen RF Tech received 510(k) clearance for an 8 channel wrist MRI coil designed to improve wrist image acquisition.

📄 Fresh Papers

• doi:10.1148/ryai.250902 - BraTS PEDs released a multi institutional MRI dataset with 457 pediatric high grade glioma cases for AI benchmarking.

• doi:10.1038/s41746-026-02670-x - Prost LM integrated MRI, PSA and reports in 3940 patients, reaching AUC 0.954 for prostate cancer diagnosis.

• doi:10.1016/j.ejro.2026.100754 - A seven site ICH study found automated AI volumetry correlated well but human correction achieved clinical grade accuracy in about 1 minute.

• doi:10.1016/j.jcmg.2026.03.008 - AI quantified breast arterial calcification in 21514 women, improving C statistic from 0.67 to 0.71 for MACE risk.

• Browse 151 new radiology AI studies from last week.

📰 Everything else in Radiology AI last week

• Aidoc raised 150 million dollars to expand its imaging AI platform, now deployed in nearly 2000 hospitals.

• A multicountry mammography study found AI CAD raised specificity from 77.0% to 88.4% and cut reading time by 32%.

• Radiology leaders emphasized post deployment monitoring to detect AI drift from scanner, protocol and population changes.

• COCA aims to detect colorectal cancer opportunistically on routine noncontrast abdominal and pelvic CT scans.

• Endometriosis imaging coverage highlighted MRI, transvaginal ultrasound and AI support for earlier noninvasive diagnosis.

• Read 15 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.