Bone tumor AI reaches AUC 0.99 in prospective testing

Good morning, there. DL-Clinic-RADS reached AUC 0.99 in prospective bone tumor testing.

I see this as a practical signal that decision support can improve risk stratification without replacing expert readers. The prospective test set matters because bone tumor assessments can change biopsy, referral, and surveillance decisions.

Where would you trust bone tumor AI in your MSK workflow?

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

• Bone tumor AI gets prospective validation

• FDA change plans are scaling in radiology AI

• CTA AI is weaker for distal occlusions

• AI finds hidden risk despite normal perfusion

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

🦴 Bone tumor AI gets prospective validation

RadAI Slice: A multicenter Radiology study gives MSK AI a stronger prospective signal.

The details:

• Included 1777 retrospective patients across 10 institutions

• Prospective test set added 152 participants from 2 centers

• DL-Clinic-RADS reached AUC 0.99 in prospective testing

• AI aid raised mean reader AUC by 0.07

• Median reading time decreased by 3 seconds overall

Key takeaway: For MSK radiology, this moves bone tumor AI closer to second-reader use, especially where Bone-RADS confidence varies.

📋 FDA change plans are scaling in radiology AI

RadAI Slice: I’m watching PCCPs because they shape how imaging AI changes after clearance.

The details:

• 1080 of 1394 FDA AI ML submissions were radiology devices

• 170 FDA devices across panels had PCCPs

• Radiology had 34 PCCP devices, 22 cleared in 2025

• Public summaries averaged 5 of 8 documentation points

• Drift triggers were absent from public summaries

Key takeaway: PCCPs can speed approved AI updates, but radiology teams still need local monitoring because public drift controls remain thin.

🧠 CTA AI is weaker for distal occlusions

RadAI Slice: This meta-analysis feels important for how we trust negative stroke AI results.

The details:

• 31 reports included 15708 patients

• ICA M1 sensitivity was 91.6% with 92.7% specificity

• M2 M3 sensitivity fell to 52.7% with 94.8% specificity

• Negative M2 M3 AI left 33.3% posttest probability at 50% pretest

Key takeaway: For stroke workflow, a negative CTA AI result should not reassure us when distal occlusion remains clinically plausible.

❤️ AI finds hidden risk despite normal perfusion

RadAI Slice: This study shows how attenuation-correction CT can add prognostic value even when myocardial perfusion is normal.

The details:

• MPI CT cohort spanned 43099 studies at 15 sites

• Analysis focused on 4552 patients with mild CAC and normal perfusion

• Proximal CAC was present in 1730 patients, 38%

• Proximal CAC raised MACE risk with adjusted HR 1.24

• Risk reclassification improved by NRI 12% for MACE and mortality

Key takeaway: Automated CTAC analysis could add prognostic information for patients with mild CAC and normal perfusion, without requiring additional imaging.

RadHarmony (24 Jul 2026)

Modality: X-ray, CT/MRI | Focus: Chest, spine | Task: Classification, detection/segmentation

• Size: 24 supported datasets. About 2.2M listed images/volumes. Patient counts are not aggregated.

• Annotations: Classification labels, segmentation masks, bounding boxes, and radiology reports.

• Institutions: Emory University, Yale University

• Availability:

  Public code and model weights: GitHub. Underlying datasets are downloaded separately under source licenses: docs.

• Highlight: Unified API harmonizes 24 radiology datasets and adds AI-assisted dataset integration.

PanDent (July 29, 2026)

Modality: X-ray (OPG) | Focus: Teeth, jaws | Task: Report generation, tooth-level diagnosis

• Size: 9,524 OPG scans/cases; patients not specified. 9,024 train and 500 test cases.

• Annotations: Expert-validated FDI tooth-level labels for 21 findings. Includes regional flags and template-style radiology reports.

• Institutions: The University of Hong Kong, The Chinese University of Hong Kong, et al.

• Availability:

  Public: Hugging Face. Code: GitHub.

• Highlight: Large OPG benchmark linking tooth-level structured findings to free-text reports for structure-language consistency.

LoMeVQA (2026-07-30)

Modality: X-ray | Focus: Chest, lungs | Task: Longitudinal VQA, visual grounding

• Size: 206K VQA samples from longitudinal CXR studies. Each sample uses 2–5 images; unique scans/patients not stated.

• Annotations: 206K longitudinal VQA pairs. Includes progress labels, free-text descriptions/reports, and bbox grounding labels.

• Institutions: Tongji University; East China Normal University

• Availability:

  Public benchmark files at GitHub; source CXR images likely restricted via MIMIC-CXR.

• Highlight: Large multi-task benchmark for temporal reasoning across serial medical images.

FAME (2026-07-30)

Modality: MRI, US, CT, OCT, CFP, endoscopy, dermoscopy, histopathology | Focus: Multi-organ; lesions/anatomy | Task: Few-shot segmentation; zero-shot segmentation

• Size: 14,958 test samples and 950 support samples from 19 public datasets. Patient count not specified.

• Annotations: Binary ROI masks. Positive and negative query cases when available.

• Institutions: Renmin University of China

• Availability:

  Unspecified. Paper: arXiv:2607.27856

• Highlight: Unified FS-MIS benchmark across 7 anatomical sites, 9 modalities, and 14 ROI classes. Includes absent-target testing and OOD shifts.

CLINLENS (28 Jul 2026)

Modality: CXR, ECG, ECHO | Focus: Chest, heart | Task: Prediction, phenotyping

• Size: 200 executable tasks. Fixed 126-task suite. Patient and scan counts not specified; 240 CXR/ECG/ECHO assets sampled for QC.

• Annotations: Reference workflows, artifact checks, cohort/temporal validation, and final numerical or categorical answers.

• Institutions: Shandong University

• Availability:

  Restricted; planned credentialed access to derived packages. arXiv

• Highlight: Links five MIMIC resources to test patient-time semantics in executable clinical analyses.

Medical-Checklist (2026-07-24)

Modality: Mixed (CT, MRI, X-ray, US) | Focus: Multi-organ, multi-system | Task: Image-text matching, VLM evaluation

• Size: 53,556 medical images; 65,464 binary tests. Patient count not reported.

• Annotations: Original positive captions. Generated negative captions with one UMLS medical term replaced. 15 concept categories.

• Institutions: Tohoku University, RIKEN AIP

• Availability:

  Unspecified; dataset and code will be public upon acceptance. Paper

• Highlight: Binary caption-choice benchmark tests whether medical VLMs reject clearly wrong medical terms.

🏛️ FDA Clearances

• K261085 - Rapid VO gained 510(k) clearance for AI triage of large vessel occlusion on CT imaging.

• K253963 - Cubresa BrainPET gained 510(k) clearance for combined brain PET and MRI imaging.

• K260206 - SkeletonPlanner gained 510(k) clearance for automated skeletal image processing and planning support.

• K260338 - ThinkSono Guidance gained 510(k) clearance for AI-guided image acquisition support.

• K260851 - Better Diagnostics Dental Assist gained 510(k) clearance for AI dental image analysis.

📄 Fresh Papers

• doi:10.1186/s12873-026-01697-3 - A prospective pediatric fracture study found AI accuracy of 95.1% but only a small recall reduction.

• doi:10.1016/j.rmed.2026.109045 - An AI-supported pulmonary nodule workflow raised ED follow-up from 53.5% to 67.9%.

• doi:10.1186/s41747-026-00775-2 - A two-center CT model for acute pancreatitis reached external AUROC 0.99 but still needs prospective testing.

• doi:10.1177/02841851261470455 - Licensed prostate MRI AI matched radiologists by AUC but missed more clinically significant cancers.

• Browse 316 new radiology AI studies from last week.

📰 Everything else in Radiology AI last week

• AI-guided handheld cardiac ultrasound in Spain cut exam cost to about €20 versus €280 for referral echo pathways.

• Qureight raised $20 million to expand AI imaging biomarkers for cardiopulmonary clinical trials.

• Octopath used biopsy-slide AI to identify rectal cancer patients more likely to benefit from irinotecan.

• SÉMIL used digital pathology and text data to stratify relapse risk in stage II bowel cancer cohorts.

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.