Radiologists prefer LLM-generated imaging histories

Good morning, there. LLMs generated histories from 28,313 patients and 20 radiologists rated them more useful.

I see this as a workflow story because better indications shape protocoling before images are acquired. The signal is not autonomy. It is safer context for radiologists who still own the interpretation.

How would you audit AI generated indications before routing exams?

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

• LLMs sharpen imaging order histories

• Neck CT misses show AI’s second look role

• Cleared mammography AI shows workload signal

• CAC CT adds AI chamber volumes for HF risk

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

Sponsored

Nalvera.AI is live: medical imaging AI in your browser, 20 jobs free

📝 LLMs sharpen imaging order histories

RadAI Slice: This Radiology study puts LLMs in a daily pain point for protocoling and interpretation.

The details:

• 28,313 UCSF patients fed a deidentified EHR dataset

• 20 radiologists reviewed 241 exams in the reader study

• Claude 3.5 Sonnet rated 68.05% factual

• The same model ranked first for protocoling in 40.87%

• Comprehensiveness drove 65.77% of overall rankings

Key takeaway: Better histories could reduce protocoling friction and focus reads, but deployment needs monitoring for omissions, drift, and accountability.

🫁 Neck CT misses show AI’s second look role

RadAI Slice: This study feels practical because the missed lesions were visible but outside the main task.

The details:

• 81,794 neck CTs yielded 123 visible lung cancers

• 80 of 123 cancers were absent from original reports

• Missed cases had 27.5 month median diagnostic delay

• Stage shift occurred in 47.5% missed vs 11.6% detected

• AI CAD found 41 of 80 missed lesions

Key takeaway: The value is not replacing the reader. It is extending vigilance at the edge of every CT field of view.

📊 Cleared mammography AI shows workload signal

RadAI Slice: This review is useful because it separates detection metrics from workflow evidence.

The details:

• 42 studies covered FDA or CE cleared mammography AI

• Digital mammography pooled AUC was 0.89

• Sensitivity was 76.3% and specificity was 89.6%

• Prospective triage studies cut workload with noninferior detection

Key takeaway: The strongest adoption case remains triage and reader support, not full replacement of radiologists.

❤️ CAC CT adds AI chamber volumes for HF risk

RadAI Slice: This study turns a familiar calcium scan into a broader cardiovascular risk marker.

The details:

• 5,892 asymptomatic patients had AI chamber volumetry

• 377 patients developed heart failure over 4 years mean follow up

• Combined model AUC was 0.80 versus PREVENT HF at 0.76

• CAC score alone reached AUC 0.70

Key takeaway: Opportunistic CT quantification may extend CAC reporting from plaque burden toward early heart failure risk stratification.

VertebralBodies (30 July 2026)

Modality: CT | Focus: Spine, vertebral bodies | Task: Segmentation, landmark localization

• Size: 1460 CT scans from TotalSegmentator and VerSe. Patient count not fully reported.

• Annotations: Thoracic and lumbar vertebral-body segmentations. Labels include T1-L6 and sacrum, with posterior elements removed.

• Institutions: LMU University Hospital Munich, German Cancer Consortium/DKFZ, et al.

• Availability:

  Public: labels and weights; two-step pipeline

• Highlight: Open vertebral-body CT labels and nnU-Net weights for precise L3 localization in body composition workflows.

LocAnyMed-200K (2026-08-04)

Modality: CT, US, X-ray, optical | Focus: Whole body, multisystem | Task: Visual grounding, object localization

• Size: 209,910 image-query-answer records. Patients/scans not specified; multiple records may share an image.

• Annotations: Bounding boxes, point coordinates, and absent-target labels. CoT rationales for 20K detection records.

• Institutions: Wuhan University, Nanyang Technological University, et al.

• Availability:

  Unspecified; code at GitHub.

• Highlight: Unifies heterogeneous medical localization datasets into free-form grounding across four modalities, with negative queries and rationale supervision.

SPARC-Rad (Unspecified)

Modality: CT, MRI, X-ray | Focus: Multi-region; normal anatomy | Task: VQA; spatial/anatomical reasoning

• Size: 300 image-question pairs from healthy-control TCIA studies. Includes 98 CT, 88 MRI, and 114 X-ray pairs. Patient count not reported.

• Annotations: Manual image-question-answer labels. Includes ground-truth answers, accepted synonyms, modality, anatomy, body region, and reasoning type. No segmentations.

• Institutions: University of Pennsylvania, University of Wisconsin–Madison, et al.

• Availability:

  Unspecified. Source images are from public TCIA.

• Highlight: Focused benchmark for spatial and anatomical reasoning in radiology VLMs, not disease diagnosis.

Spinal-Multiple-Myeloma-SEG v2 (July 2026)

Modality: DECT | Focus: Spine, lumbar vertebrae | Task: Segmentation, multimodal learning

• Size: 72 DECT exams from 67 adult patients. Trabecular masks available for 71 studies.

• Annotations: Voxel-wise trabecular bone masks for lumbar vertebrae. Existing vertebra and focal myeloma lesion masks included in v1.

• Institutions: Brno University of Technology, University Hospital Brno, et al.

• Availability:

  Public. Extension: Zenodo. Original: TCIA.

• Highlight: Adds expert-validated trabecular bone segmentation to a spectral CT multiple myeloma dataset.

🏛️ FDA Clearances

• K261349 - Airs Medical received 510(k) clearance for SwiftMR, automated software for radiology image processing.

• K253770 - GE Medical Systems SCS received 510(k) clearance for CardIQ Suite for CT image analysis workflows.

• K261535 - Hi D Imaging received 510(k) clearance for 4TAVR, software for image processing in TAVR planning.

• K260192 - ROPCA received 510(k) clearance for DIANA, AI based software for radiological image processing.

• K254096 - Velmeni received 510(k) clearance for V4D 3D software for dental 3D radiological image analysis.

• K253452 - Rivanna Medical received 510(k) clearance for Accuro 3S, an ultrasound imaging system for real time anatomy.

• Explore last week's 13 radiology AI FDA approvals.

📄 Fresh Papers

• doi:10.1148/radiol.254034 - DBT DINO used 487,975 DBT volumes for pretraining and improved density classification, but not lesion detection.

• doi:10.1016/j.acra.2026.07.036 - Automated CT peripancreatic collection volumetry achieved external Dice 0.89 and predicted organ failure with AUC 0.82.

• doi:10.1016/j.ejrad.2026.113127 - A 965 patient MRI system classified benign versus malignant small renal masses across three institutions.

• doi:10.1002/mp.70614 - A Siemens prostate MRI prototype matched radiologist PI RADS performance on 202 de novo external cases.

• Browse 266 new radiology AI studies from last week.

📰 Everything else in Radiology AI last week

• RSNA launched a knee MRI AI challenge using more than 5,000 images and reports from 16 global sites.

• Vanderbilt and Lilly are funding an EHR triage agent and future MRI safety reports for Alzheimer therapy workflows.

• ACR released a patient handout explaining AI generated radiology report summaries and why radiologists remain central.

• CT radiomics predicted MRONJ risk before therapy in 94 patients, with AUC up to 0.931 in a retrospective study.

• A prostate radiopharmaceutical therapy model estimated tissue absorbed dose from imaging in less than 23 seconds.

• Read 15 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.