Neuro AI trained on 5.24M MRI and CT volumes

Good morning, there. NeuroVFM was trained on 5.24M MRI and CT volumes from 566,915 routine clinical neuroimaging studies.

I see this as a scale signal for radiology AI because NeuroVFM learned from two decades of routine health system imaging rather than relying only on public datasets. The real test is whether its gains in report accuracy, hallucinations and critical errors hold across other hospitals and live clinical workflows.

How much local validation would you need before using a model like this?

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

• NeuroVFM trains on 5.24M brain MRI and CT volumes

• FDA clears AI for coronary CT plaque analysis

• Ultrasound AI predicts failed intussusception reduction

• Only 9 RCTs have tested imaging AI in practice

• Plus: 7 newly released datasets, 6 FDA approved devices & 4 new papers.

🧠 NeuroVFM trains on 5.24M brain MRI and CT volumes

RadAI Slice: This Nature Medicine paper puts health system scale training at the center of neuroimaging AI.

The details:

• Trained on 5.24M routine clinical MRI and CT brain volumes

• Data came from 566,915 neuroimaging studies collected over two decades

• Learned a shared representation across MRI and CT

• Generated neuroimaging reports with fewer hallucinated findings

• Reduced critical reporting errors compared with frontier models

• Paired imaging features with open source language models

Key takeaway: This is a practical signal that radiology foundation models may need health system scale clinical data, not only public imaging datasets. The next question is whether these gains generalize across institutions, scanners and reporting environments.

🫀 FDA clears AI for coronary CT plaque analysis

RadAI Slice: This clearance brings automated coronary tracing, plaque assessment and structured reporting into the syngo.via reading workflow.

The details:

• 510k cleared as a plug in for MM Reading on syngo.via

• Vessel tracing validated on 425 multi vendor CCTA datasets

• Lumen and vessel wall segmentation evaluated on a 150 case subset

• Pointwise tracing sensitivity was 95.8 percent for vessels at least 1.5 mm

• Pointwise tracing precision was 96.1 percent

• Quantifies stenosis, plaque volume and remodeling index

Key takeaway: This brings automated coronary plaque assessment directly into the reading workflow, combining quantification, reporting and treatment planning tools in one workstation environment.

🧒 Ultrasound AI predicts failed intussusception reduction

RadAI Slice: This prospective ultrasound study evaluates AI at a real pediatric treatment decision point.

The details:

• Included 5602 children from 14 tertiary hospitals

• Predicted whether air enema reduction would fail

• Prospective validation tested 190 patients

• Overall accuracy was 93.7 percent

• Senior sonographers reached 74.7 percent accuracy

• Junior sonographers reached 60.7 percent accuracy

Key takeaway: The prospective comparison against sonographers makes this more compelling than a retrospective accuracy study. The next question is whether it performs reliably during real time scanning and outside Chinese tertiary hospitals.

📊 Only 9 RCTs have tested imaging AI in practice

RadAI Slice: This review gives a useful reality check on clinical trial evidence for imaging AI.

The details:

• Scoping review found 9 randomized trials through March 2026

• AI usually worked as a clinician facing decision aid

• Sensitivity and lesion detection often improved

• Specificity remained a common limitation

• Emergency care showed smaller gains

Key takeaway: This review tempers adoption claims. I would look for multicenter RCTs with specificity, patient outcomes and workflow endpoints before scaling.

PSMA-PET-CT-Lesions (10 July 2026)

Modality: PET/CT | Focus: Whole-body, prostate | Task: Lesion segmentation, lesion detection

• Size: 597 whole-body scans from 378 male patients. 539 positive and 58 negative studies.

• Annotations: Voxel-wise manual 3D binary masks of PSMA-avid tumor lesions on PET. Primary tumors and metastases included.

• Institutions: LMU University Hospital, University Hospital Tübingen, et al.

• Availability:

  Public: TCIA PSMA-PET-CT-Lesions

• Highlight: First and largest public PSMA-PET/CT lesion dataset. Includes 18F-PSMA and 68Ga-PSMA tracers.

Expert-annotated CXR-TB (10 July 2026)

Modality: CXR | Focus: chest, lungs | Task: TB detection, abnormality classification

• Size: 1,097 CXRs from 1,097 patients; 1,039 usable after exclusions. Initial cohort: 646 TB and 451 non-TB.

• Annotations: Three B-reader reviews per CXR. Labels include normal/abnormal, TB-consistent findings, active/indeterminate TB, and microbiology-confirmed TB status.

• Institutions: Prince of Songkla University, Chiangrai Prachanukroh Hospital, et al.

• Availability:

  request-only via corresponding author; paper

• Highlight: Multi-center Thai CXR set with expert B-reader consensus and microbiological reference for external TB AI validation.

LMOD+ (June 2026)

Modality: CFP, SLO, OCT, LP, SS | Focus: Eye, retina | Task: Anatomical recognition, disease diagnosis

• Size: 32,633 ophthalmic images. Patient count not reported.

• Annotations: Bounding boxes for anatomy. Disease labels, staging labels, demographics, and free-text QA prompts.

• Institutions: Yale University, Carnegie Mellon University, et al.

• Availability:

  Public: dataset website

• Highlight: Large multimodal ophthalmology benchmark for MLLMs across 5 modalities and 12 eye conditions.

MedPMC (2026-07-08)

Modality: Mixed medical images: CT, MRI, X-ray, pathology, microscopy | Focus: Multi-organ; multi-specialty | Task: VLM pretraining; image-text retrieval

• Size: 11M image-text pairs from 6.1M PMC articles. Includes 0.6M single-panel figures, 3.1M multi-panel figures, and 7.3M subfigures. Patients not applicable.

• Annotations: Aligned captions and subcaptions. Panel-level image-text pairs, medical relevance filtering, source/license metadata.

• Institutions: Yale University, Korea University, et al.

• Availability:

  Public: Hugging Face and GitHub.

• Highlight: Continuously updatable PMC pipeline. Decomposes compound figures and aligns subcaptions. 95.3% of sampled images were medically relevant.

MFXR (2026)

Modality: X-ray | Focus: Forearm, elbow | Task: Fracture detection, anomaly classification

• Size: 4,586 X-ray images from 1,482 patients. 2,793 Monteggia fracture images and 1,793 controls.

• Annotations: Image-level diagnostic labels. Monteggia fracture vs normal or other fractures. Validated by board-certified orthopedic surgeons.

• Institutions: Macau University of Science and Technology; Shenzhen Children's Hospital; et al.

• Availability:

  Public: figshare

• Highlight: First public X-ray dataset dedicated to Monteggia fracture diagnosis. Includes benchmarks for 8 deep learning models.

MDS-Bench (2026-07-06)

Modality: Multi-modal: CT, MRI, microscopy/pathology | Focus: Multi-organ; pathology | Task: Data standardization; classification/segmentation

• Size: 1,939 target samples from 100 datasets. Patient count not reported.

• Annotations: Human-verified standardized images, per-image JSON, and dataset-level JSON. Includes source links, labels, masks, boxes, and metadata.

• Institutions: Shandong University; Stanford University

• Availability:

  Unspecified. Paper: arXiv:2607.04694

• Highlight: Benchmarks the missing upstream step: converting raw heterogeneous medical folders into VLM-ready image-text pairs.

PU2756 (04 July 2026)

Modality: US | Focus: Lung; peripheral pulmonary tumors | Task: Tumor segmentation; benign/malignant classification

• Size: 2,756 B-mode US images from 2,756 unique patients. 1,755 malignant and 1,001 benign cases.

• Annotations: Pixel-level tumor masks. Pathology-confirmed benign/malignant labels. Age bin, sex, diagnosis, and 5-fold splits.

• Institutions: The First Affiliated Hospital of Guangzhou Medical University; Zhongnan Hospital of Wuhan University; et al.

• Availability:

  Public: Figshare

• Highlight: One of the first public pulmonary US datasets with expert masks and pathology-confirmed tumor labels.

🏛️ FDA Clearances

• K260714 - Therapixel gained 510k clearance for MammoScreen 5 to aid FFDM and DBT screening mammography reads.

• K253560 - GE cleared Enhanced Boundary for PCCT to improve gray white and CSF boundary visibility on head CT.

• K260316 - Cercare cleared software for semi automatic glioma labeling, perfusion analysis and longitudinal tumor volumes.

• K253421 - Quantified Imaging cleared cloud software to generate ASL MRI relCBF and ATT perfusion maps as DICOM outputs.

• K253216 - Alphatec cleared a fluoro based intraoperative spine alignment and rod planning system with automated rod bending.

• K253921 - Silony cleared navigation instruments for spinal screw placement using CT, MR or fluoro based navigation systems.

• Explore last week's 14 radiology AI FDA approvals.

📄 Fresh Papers

• doi:10.1038/s41467-026-74985-y - Nature Communications CEUS model predicted HCC microvascular invasion using 5148 videos from 1716 patients.

• doi:10.1016/j.prro.2026.06.004 - MRI only prostate SBRT workflow used FDA cleared synthetic CT with 99.98 percent gamma pass at 3 percent 2 mm.

• doi:10.1097/RLI.0000000000001290 - Prospective study found 1.5T vessel wall MRI with DLR interchangeable with standard 3T for enhancement detection.

• doi:10.1093/neuonc/noag152 - FDA clearance testing of a brain metastasis model showed 90.0 percent sensitivity and Dice 0.86.

• Browse 307 new radiology AI studies from last week.

📰 Everything else in Radiology AI last week

• A new calculator estimates radiology AI ROI across ICH triage, PE triage and screening mammography use cases.

• AI enhanced MRI processing revealed cortical MS lesions across more than 700 ORATORIO trial participants.

• A Mayo Clinic lawsuit alleges unsafe AI deployment, data manipulation and whistleblower retaliation.

• AJR disclosures show generative AI use in radiology manuscripts rose from 1.1 percent to 7.5 percent.

• An experiment with 223 physicians found persistent overtrust in erroneous AI recommendations.

• 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.