Real-time AI boosts CT hemorrhage detection by 38%

RadAISlice Newsletter

Weekly Updates in Radiology AI

Good morning, there. Stanford’s ensemble monitoring model improved ICH detection accuracy by up to 38.6%.

I am struck by the practical advance of real-time safety monitoring for radiology AI. As adoption grows, transparent uncertainty metrics could boost trust and support physicians in managing clinical risks. This finding feels especially relevant for pathways where rapid, high-confidence triage is essential and where postmarket AI drift is an ongoing concern.

How would you use real-time AI confidence metrics in clinical CT triage?

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

• Ensemble Monitoring Model Raises Safety Bar for Radiology AI

• FDA Breakthrough Recognizes Imaging AI for Hydrocephalus Triage

• Cancer Trials: AI Syntheses and Synthetic Cohorts at ESMO

• Multicenter Deep Learning for Renal Cancer Staging

• Plus: 3 newly released datasets, 6 FDA approved devices & 4 new papers.

🩸 Ensemble Monitoring Model Raises Safety Bar for Radiology AI

RadAISlice: Stanford researchers unveiled a real-time ensemble model for monitoring radiology AI predictions during CT brain exams.

The details:

• EMM judged CT-based ICH AI predictions in real time, boosting accuracy by 38.6% for positives.

• Low false alarm rates under 1% were maintained even as ICH prevalence varied from 5% to 30%.

• Five submodels yielded best performance, especially in low-prevalence settings.

• Robustness persisted even when trained with less data, suggesting strong generalizability.

• Case-level output flagged uncertainty, guiding radiologist trust and review.

Key takeaway: Reliable, actionable uncertainty metrics will play a key role as radiology AI adoption grows. Clinicians can use real-time confidence scores to flag high-risk or low-trust predictions, helping avoid missed findings and supporting continual AI validation and safety.

🛡️ FDA Breakthrough Recognizes Imaging AI for Hydrocephalus Triage

Image from: Radiology Business

RadAISlice: Harrison.ai’s AI solution is among a minority of breakthrough devices to reach FDA clearance and Medicare eligibility.

The details:

• One of only 13% of FDA breakthrough devices to reach 510(k) marketing authorization.

• Eligible for Medicare’s NTAP, signaling high-value innovation.

• Triage/prioritization tool targets obstructive hydrocephalus using CT imaging.

• FDA breakthrough status aims to accelerate clinical access for critical use cases.

Key takeaway: Broader regulatory recognition and expedited reimbursement for imaging AI—especially in critical triage—will speed clinical validation, highlight unmet needs, and support real-world implementation in acute care pathways.

🔬 Cancer Trials: AI Syntheses and Synthetic Cohorts at ESMO

Image from: EurekAlert

RadAISlice: I’m intrigued by AI-powered synthetic data supporting multicenter oncology trial analysis and privacy protection.

The details:

• Four phase 3 trials in breast, lung, and bladder cancer presented at ESMO 2025.

• AI-generated synthetic cohorts built from 19,164 metastatic breast cancer cases.

• Synthetic datasets aim to enhance collaboration and preserve patient privacy.

• Blood biomarker (KIM-1) also highlighted as a kidney cancer guide.

Key takeaway: AI-synthesized cohorts are paving the way for scalable, privacy-preserving datasets in multicenter radiology and oncology research—critical for robust trial design and data sharing.

🦠 Multicenter Deep Learning for Renal Cancer Staging

RadAISlice: A multicenter study demonstrated generalizable CT-based AI for pre-op clear cell RCC staging using external datasets.

The details:

• 1,148 patients from 5 sites, with two external validation datasets.

• Micro-AUCs of 0.939–0.954 for T staging in external cohorts.

• Collaboration with radiologists boosted diagnostic accuracy.

• Interpretability enhanced through Grad-CAM heatmaps.

Key takeaway: Externally validated, interpretable AI staging shows promise for standardizing complex renal cancer assessment across institutions, supporting radiology-pathologist teamwork in clinical workflows.

LUS-BALD (2025)

Modality: Ultrasound | Focus: Lung | Task: Detection, Localization

• Size: 401 images from 152 patients

• Annotations: Polygonal bounding boxes of LUS vertical artifacts, YOLO-style text files

• Institutions: Makerere University, Mulago National Referral Hospital

• Availability:

  Public (https://doi.org/10.17632/h923m366sf.2)

• Highlight: First public dataset with polygonal box annotations for vertical artifacts in lung ultrasound

MultiTIPS (2025-10-12)

Modality: CT | Focus: Liver, Portal vein | Task: Survival prediction, Complication prediction

• Size: 382 patients with preoperative multiphase CT, clinical features, and follow-up; portal vein segmentation masks for all, voxel-level annotation for 32 cases

• Annotations: Segmentations of portal vein (central/peripheral), liver, inferior vena cava; structured clinical data; treatment outcomes (survival, OHE, PPG, and more)

• Institutions: Beijing University of Posts and Telecommunications, Nanfang Hospital, et al.

• Availability:

  public (https://github.com/djh-dzxw/TIPS_master)

• Highlight: First public multi-center TIPS prognosis dataset with multimodal data and expert segmentations, enabling survival and multi-task outcome prediction; supports semi-supervised and foundation model segmentation pipelines.

LaryngealCT (2024-07)

Modality: CT | Focus: Larynx, Head and Neck | Task: Classification, Segmentation

• Size: 1,029 scans, 1,029 patients

• Annotations: T-stage classification (Tis-T4), clinical metadata, CT volumes with laryngeal ROI

• Institutions: Deakin University, Manipal Academy of Higher Education

• Availability:

  Request-only (TCIA license) + public code/metadata: https://github.com/funzi-son/LaryngealCT

• Highlight: First benchmark laryngeal cancer CT dataset with reproducible ROI crops, expert validation, code, and metadata for AI model development.

🏛️ FDA Clearances

• K251602 - Alphenix (Canon), FDA-cleared, delivers real-time X-ray fluoroscopy for interventional radiology procedures.

• K251651 - Philips EPIQ and Affiniti Series, FDA-cleared, advanced diagnostic ultrasound for multipurpose clinical imaging.

• K250060 - GENORAY GT300/GT300-C, dental CT, FDA clearance for high-res 3D dental and jaw imaging.

• K252911 - iRay Lux HD 2530 Detector, digital flat-panel X-ray, FDA-c for radiology image acquisition.

• K252851 - deepLive (DAMAE Medical), optical coherence tomography system, FDA-cleared for clinical microstructure imaging.

• K252214 - AIAS Cephalon (metamorphosis GmbH), FDA-cleared AI software for automated radiology image processing.

• Explore last week's 7 radiology AI FDA approvals.

📄 Fresh Papers

• doi:10.1136/jnis-2024-022254 - Multicenter trial: AI-assisted LVO detection on CT angiography raised resident accuracy (AI AUROC 0.973; +4% sensitivity with AI support).

• doi:10.1038/s41746-025-01986-4 - UMORSS, a multicenter, uncertainty-aware ultrasound AI tool, boosted radiologist AUC (+10.6%) and sensitivity (+22.5%) for ovarian cancer risk assessment.

• doi:10.1007/s00330-025-12052-8 - Dual-site knee MRI study: deep learning AI raised resident accuracy and sensitivity (10% faster reads, high AUC for 8+ MSK pathologies).

• doi:10.1016/j.acra.2025.09.048 - Meta-analysis: MRI deep learning for colorectal cancer lymph node metastasis (n=1,850) outperformed radiologists in sensitivity and AUC.

• Browse 232 new radiology AI studies from last week.

📰 Everything else in Radiology AI last week

• Lunit acquired Prognosia to accelerate market entry of its 5-year breast cancer risk AI, which outperformed standard methods and received FDA Breakthrough Device Designation.

• The JAMA Summit report synthesizes expert policy, clinical, and regulatory perspectives on healthcare AI adoption and practical implementation for clinicians.

• AI predicts pediatric growth spurts from 5,700+ cephalometric X-rays, reducing prediction failures by 67% and halving annotation burden.

• Read 6 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. I’m building a reader-curated list of radiology AI tools. which one would you nominate?