96.6% staging accuracy for hip osteonecrosis AI

Good morning, there. OsteoAID reached 96.6% external staging accuracy for hip osteonecrosis on xray.

I see this as a meaningful test of whether xray AI can extend specialist pattern recognition. The result feels relevant for sites where MRI access or hip preservation expertise is limited.

Would you use xray AI to guide earlier hip preservation decisions?

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

• Xray AI stages hip osteonecrosis at specialist level

• AI triage accelerates pulmonary embolism response

• LLM catches report text errors across centers

• Local audits expose nodule AI workflow limits

• Plus: 2 newly released datasets, 1 FDA approved device & 4 new papers.

🦴 Xray AI stages hip osteonecrosis at specialist level

RadAISlice: This study stands out for scale, external testing, and direct comparison with clinicians.

The details:

• 16,399 patients and 20,722 hip xrays from 6 Chinese hospitals

• Internal staging accuracy rose from 87.4% with YOLOv8 to 96.4%

• External staging accuracy was 96.6%, similar to expert orthopaedists

• AI assistance improved radiologist accuracy by 21.5 percentage points

• Challenging image set accuracy remained 93.8%

Key takeaway: This could make ONFH triage more accessible where MRI or specialist reads are scarce, but prospective deployment data still matter.

🚑 AI triage accelerates pulmonary embolism response

RadAISlice: This is a practical deployment story with measurable effects on acute CT workflow.

The details:

• 424 symptomatic central or segmental PE patients met inclusion criteria

• PERT activation increased from 9.2% to 84.5% after AI implementation

• CT completion to PE diagnosis fell from 2.1 to 1.3 hours

• Diagnosis to interventional therapy was reduced by 50%

• Mortality and length of stay trends did not reach significance

Key takeaway: The workflow signal is strong, but outcome benefit still needs prospective validation before calling it practice changing.

📝 LLM catches report text errors across centers

RadAISlice: This report QA study feels unusually close to daily radiology operations.

The details:

• 95,582 reports from center 1 were reviewed across xray, CT, and MRI

• Confirmed internal text errors appeared in 4.6% of reports

• Qwen3 reached 0.99 sensitivity and 0.99 specificity

• External center balanced accuracy was 0.94

• Mean review time was 5.8 seconds versus 55.8 to 67.8 for radiologists

Key takeaway: LLM QA may first prove useful as a safety net for explicit report inconsistencies, not as image interpretation.

🔍 Local audits expose nodule AI workflow limits

RadAISlice: This audit is a useful reminder that postmarket performance can diverge from benchmark promise.

The details:

• 13,536 consecutive chest CT exams were analyzed after deployment

• AI marked 67.5% positive, while reports documented nodules in 29.0%

• Review adjusted sensitivity was 81.2% and specificity was 38.2%

• Positive predictive value was 35.4% after image review

• Inpatient and ED exams contributed 76.1% of discordance

Key takeaway: Before scaling detection AI, local indication specific audits should be part of governance, not an optional afterthought.

DoseRAD2026 (April 2026)

Modality: CT, MRI | Focus: Thorax, abdomen | Task: Dose calculation, dose prediction

• Size: 115 paired CT-MRI patients. 75 training and 40 testing cases.

• Annotations: Beam-level photon and proton MC dose maps. JSON beam configurations are included.

• Institutions: LMU University Hospital, German Cancer Research Center (DKFZ), et al.

• Availability:

  Public training set on Zenodo. Test set withheld until March 2030.

• Highlight: First public paired CT-MRI benchmark with beam-level photon and proton MC dose ground truth.

MultiCaRe-Thorax (5 Oct 2026)

Modality: X-ray | Focus: Thorax, chest | Task: Multi-label disease classification; reproducible NLP labeling

• Size: 5,252 CXR images from 3,415 patients.

• Annotations: 16 binary thoracic disease labels from negation-aware NLP. Validated on 500 reports.

• Institutions: The Hashemite University; Higher Colleges of Technology

• Availability:

  Public: Zenodo, GitHub

• Highlight: First curated thorax benchmark from MultiCaRe with a reproducible negation-aware labeling pipeline.

🏛️ FDA Clearances

• K260391 - Coreline Soft received 510k clearance for AVIEW Aorta Metrics, CT software for automated aortic analysis.

📄 Fresh Papers

• doi:10.1007/s11604-026-02103-5 - Prospective ultrafast stroke MRI cut acquisition time by 61.7% while remaining interchangeable for lesion detection.

• doi:10.1002/jmri.70599 - A 1,408 patient multicenter JMRI study found 2.5D quantitative MRI best predicted glioma IDH and 1p19q status.

• doi:10.1016/j.acra.2026.09.071 - A hybrid ultrasound model improved O RADS 4 adnexal mass risk stratification and reader agreement in validation.

• doi:10.1186/s13244-026-02443-w - A dual center CT deep learning model predicted acute pancreatitis severity with external AUC 0.910.

• Browse 340 new radiology AI studies from last week.

📰 Everything else in Radiology AI last week

• Stanford reported a local LLM with 99.7% sensitivity for critical pediatric radiology result detection.

• University of Washington found ICH CT triage cut median positive report turnaround from 214 to 201 minutes.

• Pennsylvania House Bill 1925 would require imaging centers to disclose AI use and review performance yearly.

• UVA won a $4M NIH grant to extract DENSE quality strain from routine cardiac cine MRI across centers.

• SIIM CAIMI leaders emphasized that imaging AI is moving from evaluation toward workflow implementation.

• Read 10 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.