$16M US trial tests AI in breast cancer screening

RadAI Slice Newsletter

Weekly Updates in Radiology AI

Good morning, there. The $16M PRISM trial will randomize AI in breast mammography for 6 states.

This is the first large multicenter randomized US trial assessing AI in breast cancer screening. Its design and outcomes will inform both clinical implementation and payer policy decisions for AI in radiology. Results could define standards for real-world AI use in mammography nationwide.

PS: If this touches your work, hit Reply with one note. I read every message.

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

• Landmark $16M AI Trial in US Breast Cancer Screening Launches

• Cigna Expands Nationwide Coverage for CT Imaging AI Tools

• Multicenter AUC 0.87 Model Stratifies HCC Risk in Cirrhosis

• AI-Generated Chest X-Ray Reports Nearly Match Radiologist Acceptability

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

🧬 Landmark $16M AI Trial in US Breast Cancer Screening Launches

Image from: EurekAlert

RadAI Slice: A nationwide, randomized US mammography trial will prospectively measure AI value in breast cancer screening.

The details:

• $16M PCORI-funded, six-state, multicenter randomized trial

• Hundreds of thousands of mammograms, 7 academic centers inc. UCLA, UM

• ScreenPoint Transpara AI fully integrated; Aidoc aiOS for workflow

• Randomized to AI-assisted or radiologist-only groups with real-world outcomes

• Outcomes: Cancer detection rate, recall, radiologist/patient trust in AI

Key takeaway: This trial sets a clinical and policy benchmark for US breast imaging AI. Real-world data could drive payer reimbursement and large-scale workflow adoption.

💡 Cigna Expands Nationwide Coverage for CT Imaging AI Tools

Image from: Radiology Business

RadAI Slice: Cigna will begin reimbursing CT plaque analysis AI nationwide across commercial and Medicare plans.

The details:

• Effective October 1; >16M Cigna members will be covered

• Covers AI-based plaque analysis inc. HeartFlow and others

• Follows UnitedHealthcare’s coverage expansion in early 2024

• Change driven by EviCore radiology benefits update in July

Key takeaway: Major payer coverage gives CT plaque AI tools access to tens of millions, signaling payer acceptance and clinical adoption of AI in routine cardiology imaging.

🧪 Multicenter AUC 0.87 Model Stratifies HCC Risk in Cirrhosis

RadAI Slice: Multimodal CT radiomics+deep learning model outperforms clinical scores for HCC risk in cirrhosis.

The details:

• N=2,411 across 7 Chinese centers; median FU 43 months

• Outperforms standard aMAP; AUC 0.81–0.87 (3 cohorts)

• High-risk: 26.3% 3y HCC incidence vs 1.7% for low-risk

• Stepwise model (aMAP → aMAP-CT) improves stratification

Key takeaway: Validated multicenter CT-based tools can personalize surveillance for cirrhosis, demonstrating generalizable methods for radiology-driven precision risk models.

🖥️ AI-Generated Chest X-Ray Reports Nearly Match Radiologist Acceptability

RadAI Slice: Large-scale, multicenter AI model delivers chest x-ray reports nearly matching clinical acceptability of radiologists.

The details:

• 8.8 million CXRs, 42 sites, multicountry training

• Acceptability: AI 88.4% vs radiologists 89.2% (p=0.36)

• More stringent (no revision): AI 66.8% v rads 75.7% (p<0.001)

• AI has higher sensitivity, lower specificity for abnormalities

Key takeaway: AI-generated chest x-ray reports approach clinical standards for initial reads, though current limitations mean continued human oversight remains essential.

MultiD4CAD (2025)

Modality: CT | Focus: heart, coronary vessels | Task: segmentation, classification

• Size: 118 patients, 118 CCTA scans

• Annotations: Epicardial and pericoronary adipose tissue masks, CAD labels, clinical data

• Institutions: University of Palermo, ICAR-CNR et al.

• Availability:

  restricted (Zenodo link)

• Highlight: Multimodal dataset including CCTA, tissue segmentations, and rich clinical risk factors for CAD.

Ultrasound Spinal Cord Dataset (USSC) (2025-08-13)

Modality: Ultrasound | Focus: Spinal cord | Task: Injury localization, anatomical segmentation

• Size: 10,223 porcine scans (N=25), 86 human scans (N=8)

• Annotations: Bounding boxes for injury, pixel-level masks for dura, CSF, pia, spinal cord, hematoma

• Institutions: Johns Hopkins University, Cleveland Clinic

• Availability:

  Public GitHub

• Highlight: Largest open spinal cord ultrasound dataset; includes healthy/injured, multi-class labels, benchmarks DL models.

MedForensics (2025-09-19)

Modality: CT, MRI, X-ray, Ultrasound, Endoscopy, Histopathology | Focus: breast, brain | Task: deepfake detection, image classification

• Size: 116,000 images (real + synthetic), from multiple modalities; roughly 58,000 real and 58,000 synthetic

• Annotations: real/fake image labels; linked to modality and generating model, no pixel-level segmentations

• Institutions: The Hong Kong University of Science and Technology (Guangzhou), The Hong Kong University of Science and Technology

• Availability:

  public (upon publication) arXiv link

• Highlight: First large-scale medical AI-deepfake dataset spanning six modalities and 12 generation models

🏛️ FDA Clearances

• K251386 - Fujifilm’s ECHELON Synergy MRI system approved for high-definition multiplanar imaging; supports radiology diagnosis and monitoring.

• K251167 - Shanghai United Imaging’s uDR Aurora CX, a new stationary x-ray system, cleared for general radiology imaging; compatible with AI workflow.

• K250883 - Olympus ultrasonic probes UM-3R and UM-G20-29R cleared; support high-resolution diagnostic ultrasound imaging.

• K250369 - Axial3D Insight receives 510(k): automates radiological image processing, supporting complex anatomical analysis.

• Explore last week's 6 radiology AI FDA approvals.

📄 Fresh Papers

• doi:10.1109/JBHI.2025.3611004 - A fuzzy label/active learning approach improves intestinal ulcer segmentation model robustness and cross-dataset generalizability for IBD applications.

• doi:10.1097/MD.0000000000044493 - Meta-analysis of >20 AI studies in MS diagnosis (MRI-based) finds pooled sensitivity 0.93, specificity 0.95, highlighting strong AI contributions in clinical differentiation.

• doi:10.1007/s00330-025-12029-7 - Yonsei University shows deep learning MRI reconstruction for TMJ halves scan time with equivalent diagnostic performance and less noise.

• doi:10.1016/j.acra.2025.09.009 - Large US chest CT study validates deep learning–based CAC scoring with strong agreement (ICC=0.987) across various protocols and COVID-19 cohorts.

• Browse 215 new radiology AI studies from last week.

📰 Everything else in Radiology AI last week

• JAMA analysis finds inconsistent FDA safety and efficacy reporting for cleared AI/ML-enabled devices, calling for post-market surveillance and AI-specific guidelines.

• CMS pilots AI for Medicare prior authorization in 6 states and launches the $16M PRISM trial for breast cancer screening AI, highlighting regulatory momentum for radiology AI adoption.

• MD Anderson reviews imaging AI advances in early detection, personalized oncology, and proton therapy with supporting phase III data, ahead of ASTRO 2025.

• MSK Cancer Center presents phase II data: proton therapy doubles survival for leptomeningeal metastasis; also spotlights LLMs for pain note summarization.

• Researchers urge stricter US transparency for financial ties between AI radiology device makers and radiologists, citing regulatory loopholes.

• Read 13 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.