AI flags CRC on routine CT with 99.8% specificity

Good morning, there. COCA detected CRC on noncontrast CT with 86.6% sensitivity and 99.8% specificity.

I see this as a practical test of opportunistic detection on scans we already read. The 27,433 patient real world cohorts make the workflow signal harder to ignore.

How would opportunistic CRC detection fit into your CT workflow?

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

• CT opportunistic screening flags CRC at scale

• Chest CT model quantifies ARDS in one pass

• 🩻 Chest X ray triage gets a silent trial

• Oncology AI still lacks prospective evidence

• Plus: 2 newly released datasets, 6 FDA approved devices & 4 new papers.

🎯 CT opportunistic screening flags CRC at scale

RadAISlice: I see this as one of the clearest examples yet of opportunistic AI screening from routine CT.

The details:

• Development used 2678 patients from 2 centers

• External validation covered 2053 patients across 6 centers

• Real world cohorts included 27,433 consecutive patients

• COCA reached 86.6% sensitivity and 99.8% specificity in cohort 2

• Reader study showed 20.4% sensitivity gain with AI support

Key takeaway: This could move CRC detection upstream by turning routine CT into an opportunistic safety net, if sites can manage alerts and follow up.

🫁 Chest CT model quantifies ARDS in one pass

RadAISlice: This feels clinically relevant because it ties chest CT directly to ICU decision support.

The details:

• Trained on more than 50000 CT volumes

• Validated across 6153 individuals from 6 centers

• ARDS diagnosis AUC reached 0.87 and respiratory failure AUC 0.97

• CT estimated P to F ratio with correlation of 0.83

Key takeaway: For radiologists supporting critical care, this points toward CT reports that include reproducible severity and prognosis signals rather than description alone.

🩻 Chest X ray triage gets a silent trial

RadAISlice: This prospective NHS study gives us unusually practical evidence for normal CXR triage.

The details:

• 63,083 adult CXRs were analyzed across 5 NHS sites

• AI labeled 20% as normal and 80% as abnormal

• Sensitivity was 97%, specificity 35%, and NPV 94%

• Clinically significant miss rate was estimated at 0.05%

• Normal report concordance could deprioritize 18.5% of CXRs

Key takeaway: The result supports cautious normal CXR triage to protect turnaround time while keeping a visible failure analysis process.

📋 Oncology AI still lacks prospective evidence

RadAISlice: This review is a useful reality check for imaging AI approvals in cancer care.

The details:

• 149 of 1,008 FDA AI devices had oncology indications

• Radiology accounted for 46% and radiation oncology 38%

• 76% reported clinical testing with patient data

• Only 21% had clinician in loop testing

• Only 5% described prospective testing

Key takeaway: For adoption committees, the paper argues evidence expectations should rise with decision risk, especially for CAD tools.

UTSW-Glioma (2026)

Modality: MRI | Focus: brain, glioma | Task: segmentation, molecular profiling

• Size: 625 patients with preoperative multi-contrast MRI; 362 cases also have high-quality manual segmentations

• Annotations: Tumor subregion masks for ET, NCR, and ED. Metadata include IDH, 1p19q, MGMT, tumor type, and grade.

• Institutions: University of Texas Southwestern Medical Center, Mayo Clinic

• Availability:

  public (TCIA)

• Highlight: Large public glioma MRI dataset linking four MRI contrasts with molecular markers and expert-refined tumor segmentations.

ISLES'24 (2026-04-07)

Modality: CT/MRI | Focus: brain, cerebrovascular | Task: lesion segmentation, outcome prediction

• Size: 245 patients; 149 public training cases and 96 hidden test cases. Multimodal hyperacute CT and follow-up acute MRI.

• Annotations: Manual CTA vessel occlusion masks, final infarct segmentations on DWI with human-AI refinement, plus Circle of Willis pseudo-label segmentations and clinical outcomes to 3 months.

• Institutions: Technical University of Munich, University Hospital Zurich et al.

• Availability:

  public training set via Zenodo; hidden test set via Grand Challenge

• Highlight: Longitudinal stroke dataset linking hyperacute CT, post-reperfusion MRI, and 3-month outcomes for modeling infarct evolution.

🏛️ FDA Clearances

• K252548 - Siemens AI Rad Companion Organs RT adds organ segmentation support for radiation therapy planning workflows.

• K253930 - Overjet Iris received FDA clearance for AI image enhancement and analysis support in dental radiography workflows.

• K253689 - Siemens syngo Dynamics VA41F adds automated image processing support within cardiovascular imaging workflows.

• K254021 - Rivanna Accuro XV is an ultrasound system that supports real time visualization for procedural and diagnostic use.

• K253862 - Fujifilm APERTO Lucent MRI System received FDA clearance for high quality MR imaging in routine clinical practice.

• K253446 - Dk Medical Systems AeroDR TX c02 is a stationary x ray system cleared for routine diagnostic imaging.

• Explore last week's 11 radiology AI FDA approvals.

📄 Fresh Papers

• doi:10.1038/s43018-026-01147-w - A multireader multicase trial showed an AI assistant improved junior radiologist performance for incidental pulmonary nodules on CT.

• doi:10.1093/neuonc/noag088 - A prospective biopsy controlled trial validated MRI AI mapping of glioblastoma infiltration with 0.84 AUC and survival relevance.

• doi:10.1016/j.acra.2026.03.058 - A multinational multireader mammography study found AI CAD improved average AUC from 0.799 to 0.851 and cut reading time from 121.5 to 83.2 seconds.

• doi:10.1186/s13244-026-02286-5 - A prospective ultrasound model predicted carotid plaque response to statins within 6 months, reaching 93.7% specificity at the key clinical time point.

• Browse 158 new radiology AI studies from last week.

📰 Everything else in Radiology AI last week

• GE HealthCare and RadNet expanded their AI mammography partnership globally, adding lesion localization, density, and safeguard review tools.

• A 215381 study ICU chest x ray dataset with structured labels and severity grading is now openly available for clinically relevant model development.

• Radiologists preferred a domain specific CT reporting model over general LLMs after reviewing 200 reports at Moffitt Cancer Center.

• A healthcare operations analysis found 80% of systems struggle to measure AI ROI, highlighting governance and workflow gaps.

• Radiology received $90.4 million in industry research payments in 2024, only 1.1% of total U.S. research funding.

• 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.