Explainable AI improved head CT decisions OR 1.84

Good morning, there. Correct explainable AI improved head CT decisions with OR 1.84 in a 125 reader study.

I see this as a useful reminder that AI value depends on how readers use advice. The study separates underreliance from overreliance, which feels closer to real PACS behavior than accuracy alone.

How should radiology teams train readers to calibrate trust in AI advice?

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

• Trust calibration in head CT AI

• Breast US AI and benign biopsy reduction

• Radiology societies press FDA on gen AI safety

• Agentic AI scores glioma follow up MRI

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

🧠 Trust calibration in head CT AI

RadAISlice: This study stands out because it measures reliance, not just model accuracy.

The details:

• 50 radiologists and 75 nonradiologist clinicians read head CT cases

• Each session used 50 scans with 80% correct and 20% incorrect AI advice

• Correct explainable advice improved performance with OR 1.84

• Underreliance on correct advice harmed accuracy more than overreliance

• Explainable AI did not significantly shorten reading time

Key takeaway: The result makes AI training a workflow issue: radiologists need cues that improve calibration, not just saliency maps.

🩺 Breast US AI and benign biopsy reduction

RadAISlice: This is a clinically useful signal, but it also shows why AI downgrades need caution.

The details:

• 271 BI RADS 4 lesions from 253 women were assessed after implementation

• Malignancy rate was 23.2%, with 63 cancers and 208 benign results

• AI labeled 60.6% of benign lesions as BI RADS 2 or 3

• AI labeled 90.5% of malignant lesions as BI RADS 4

• 6 of 63 cancers were assigned BI RADS 2 or 3 by AI

Key takeaway: This supports biopsy triage research, but not autonomous downgrading; clinical history and reader judgment remain essential.

⚖️ Radiology societies press FDA on gen AI safety

Image from: Radiology Business

RadAISlice: ACR and RSNA focused on monitoring, accountability, and patient safety.

The details:

• FDA received more than 300 public responses on generative AI regulation

• ACR and RSNA both submitted imaging focused comments

• RSNA urged continuous real world monitoring after deployment

• Societies said manufacturers should retain safety responsibility

Key takeaway: The comments could shape postmarket surveillance and accountability for imaging AI, especially adaptive and generative tools.

🤖 Agentic AI scores glioma follow up MRI

RadAISlice: A Duke team tested a multiagent system for standardized BT RADS classification.

The details:

• 492 posttreatment glioma MRI exams were evaluable

• System accuracy was 76.0% versus 57.5% for initial assessments

• Cohen kappa versus expert reference was 0.71

• BT 4 positive predictive value reached 92.9%

• Local open weight inference was framed as a privacy advantage

Key takeaway: This is a plausible second read use case for structured neuro oncology follow up, but it still needs external and prospective testing.

RetinaGroundBench (2026-10-22)

Modality: OCT | Focus: Retina, macula | Task: Visual grounding, VQA

• Size: 815 B-scans; patient count not specified. Images come from OCTDL and OLIVES.

• Annotations: 4,537 expert bounding boxes for 13 retinal biomarkers across 8 diseases.

• Institutions: Medical University of Vienna

• Availability:

  Public: GitHub

• Highlight: First retinal OCT benchmark for text-prompted biomarker grounding with MLLMs.

ViBMR (2026-10-21)

Modality: MRI | Focus: Brain | Task: Radiology report generation; VLM benchmarking

• Size: 3,649 patients/studies and 11,875 MRI series. Includes T1, T2, FLAIR, and T1c.

• Annotations: Routine Vietnamese clinical reports. Findings sections are used as text targets. No segmentation masks.

• Institutions: Hanoi University of Science and Technology; Télécom SudParis et al.

• Availability:

  Restricted; planned non-commercial research release. No link specified.

• Highlight: First large-scale multi-sequence brain MRI dataset paired with native Vietnamese routine reports.

PediaDeviceSeg-1K (2026-10-20; public release planned upon publication)

Modality: X-ray | Focus: Pediatric chest; lungs | Task: Device segmentation; device-aware CXR analysis

• Size: 1,000 frontal/AP CXRs from 100 pediatric CICU patients. Selected from a 3,122-image source cohort.

• Annotations: Fine-grained ICU device labels. Polylines or polygons for tubes, lines, wires, catheters, leads, and other hardware.

• Institutions: University of Notre Dame; Boston Children's Hospital; et al.

• Availability:

  Unspecified/request-only during peer review. Public repository planned upon publication; no link provided.

• Highlight: First public pediatric CICU CXR dataset focused on clinically grounded device annotations.

🏛️ FDA Clearances

• K262301 - Rapid ICH v4.0 received 510k clearance for CT triage and prioritization of intracranial hemorrhage.

• K262134 - NeuroDR received 510k clearance for AI assisted neuro MRI detection and radiologist review support.

• K261733 - Ligence Heart received 510k clearance for AI cardiac ultrasound segmentation and analysis support.

• K260523 - QGuideCT 1.0 received 510k clearance for automated CT image processing and clinician interpretation support.

📄 Fresh Papers

• doi:10.1148/ryai.260538 - TRAM AI offers a 14 item checklist for judging whether public imaging datasets are usable for AI development.

• doi:10.1007/s00330-026-12969-8 - A prospective chest radiography reader study found night shift fatigue lowered accuracy, and AI did not restore performance.

• doi:10.1016/j.acra.2026.09.075 - A review of regulator cleared acute stroke CT AI found sociodemographic subgroup performance reported in only 1 study.

• doi:10.1111/bju.70491 - MRUSuNet combined MRI and TRUS for prostate cancer localization, reaching sensitivity 0.80 across test cohorts.

• Browse 230 new radiology AI studies from last week.

📰 Everything else in Radiology AI last week

• Elucid gained FDA clearance for AI FFR CT from CCTA, becoming a third US commercial FFR CT vendor.

• SimBioSys received FDA clearance for MRI based breast cancer recurrence risk scoring after a 2129 patient study.

• DROID RV trained on over 380000 echo videos and linked AI RV dilation to twice the 5 year mortality risk.

• Penn Medicine described Percival, an open source CT model trained on about 400000 scans for opportunistic risk detection.

• LLMs recommended unnecessary care in 5.4% of scenarios, but thyroid ultrasound overuse reached 42.8%.

• Read 9 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.