RadAI Slice Newsletter Weekly Updates in Radiology AI |
Good morning, there. AI-CAD achieved 89.9% sensitivity and 94.3% specificity in 24,543 mammograms. I see this as a concrete demonstration of AI supporting radiologists in large-scale breast cancer screening using real-world data. This finding feels especially relevant as it highlights strengths and ongoing needs, such as in dense breasts and lack of prior exams. Integrating AI into screening workflows may help optimize cancer detection and patient outcomes. How would you integrate AI-CAD tools into your breast imaging workflow?
Here's what you need to know about Radiology AI last week: AI-CAD Enhances Breast Cancer Detection in Screening Practice New Open-Source Federated Learning Framework for Radiology AI AI on Chest X-rays Outperforms Standard COPD Mortality Stratification Dual-Center MRI Study: Integrated Model Accurately Predicts Cervical Cancer Invasion Plus: 5 newly released datasets, 5 FDA approved devices & 4 new papers.
|
Which AI tool do you use most often in your daily work? |
🔍 AI-CAD Enhances Breast Cancer Detection in Screening Practice RadAI Slice: A large prospective study confirms high sensitivity and workflow value of AI-CAD in mammographic screening. The details: 24,543 women screened with Lunit Insight AI-CAD; 2021–2022 period AI-CAD: 89.9% sensitivity, 94.3% specificity, PPV 8.7%, exceeding BI-RADS thresholds AI-CAD found 3.4% of cancers radiologists missed, but missed 8.1% found by radiologists False negatives more common with dense breasts; AI used no prior exams False negative rate overall: 10.1% for AI-CAD in this cohort
Key takeaway: AI-CAD can improve screening outcomes but has limits, especially without prior exams or in dense breast tissue. These results will guide safe, effective adoption to support radiologist performance. |
🤝 New Open-Source Federated Learning Framework for Radiology AI RadAI Slice: A new federated learning framework lets centers collaborate across architectures and data, enhancing robust AI for imaging. The details: UnifiedFL works even when clients use CNN, GNN, or MLP in the same collaboration Proven on MedMNIST classification and hippocampus segmentation benchmarks Open code; strong performance on multi-institutional imaging tasks Dynamic aggregation; promotes diversity and generalization
Key takeaway: We finally see practical, open-source tools for federated, privacy-preserving AI development in radiology. This approach supports robust, cross-institutional model deployment without centralizing data. |
🫁 AI on Chest X-rays Outperforms Standard COPD Mortality Stratification RadAI Slice: AI algorithms using chest X-rays deliver superior 10-year mortality risk prediction in COPD versus GOLD grading. The details: Analyzed 4,226 x-rays from mild to severe COPD patients; external Asian validation 16% increase in respiratory mortality per 5-yr AI risk increment (adjusted) AI AUC: 0.76 for 10-year mortality versus 0.61 for GOLD–p<0.001 Pulmonary function decreased as AI scores rose
Key takeaway: AI can noninvasively stratify COPD risk from standard imaging; integrating these scores could reshape COPD management and follow-up strategies in radiology. |
📈 Dual-Center MRI Study: Integrated Model Accurately Predicts Cervical Cancer Invasion RadAI Slice: Dual-center study shows multimodal MRI-based model outperforms radiomics or deep learning alone for cervical cancer spread. The details: 290 patients from 2 centers with pre-op MRI Combined radiomics, habitat imaging, and deep learning—AUC 0.91 external test Model outperformed individual modalities and clinical assessment Calibration and decision analysis confirmed clinical utility
Key takeaway: Blending radiomics, habitat imaging, and advanced ML offers highly accurate, generalizable pre-surgical decision support and could streamline cervical cancer care. |
MORE (2025-10-30) Modality: CT | Focus: multi-organ | Task: reconstruction, generalization Size: 65,755 CT slices from 135 patients Annotations: Lesion labels and anatomical region (9 body regions, 15 lesion types); 2D DICOM slices with PNGs Institutions: Shanghai Jiao Tong University, Suzhou Xiangcheng People’s Hospital Availability: Highlight: First public multi-organ CT reconstruction benchmark; covers 9 anatomies and 15 lesion types for robust generalization.
|
BHD (2025-10-23) Modality: MRI, CT | Focus: Brain | Task: Classification, Prognosis Size: MRI: 10,709 dementia cases + 10,709 controls; CT: 57,242 dementia cases + 57,242 controls; over 830,000 people Annotations: Linked EHRs, dementia labels (ICD10/medications), imaging sequence, NLP-extracted brain phenotypes; 713 scans manually annotated for validation Institutions: University of Edinburgh, University of Dundee et al. Availability: Highlight: Largest national clinical brain imaging cohort with linked EHR and radiology reports, curated for real-world AI research in dementia.
|
SEG.A. (Segmentation of the Aorta Challenge Dataset) (2023-05-15) Modality: CT | Focus: Aorta, vessel tree | Task: Segmentation, surface meshing Size: 56 training scans (anonymized, multicenter), 56 patients; test set: 157 scans (hidden) Annotations: Manual expert-validated masks segmenting the aortic vessel tree Institutions: Graz University of Technology, University Medicine Essen et al. Availability: Highlight: First large, multicenter, public CTA dataset for aortic vessel tree segmentation; includes diverse pathologies and mesh-ready ground-truth.
|
CoSpine (2025) Modality: fMRI | Focus: brain, cervical spinal cord | Task: task activation, AI methods development Size: 61 participants (39 pain, 22 motor); whole-brain and spinal cord fMRI scans Annotations: Task event timing, behavioral ratings, physiological recordings, segmentations (brain and spinal cord), field maps Institutions: Institute of Psychology Chinese Academy of Sciences, Shandong Provincial Hospital et al. Availability: Highlight: First open, BIDS-compliant, simultaneous cortico-spinal fMRI dataset with raw and preprocessed data for pain and motor tasks
|
CSpineSeg (2025) Modality: MRI | Focus: Cervical spine | Task: Segmentation, anatomical labeling Size: 1,255 MRI exams, 1,232 patients Annotations: Manual segmentations of vertebral bodies and discs for 481 patients; rest weakly labeled Institutions: Duke University, Gannon University Availability: Highlight: Largest public cervical spine MRI segmentation set with manual anatomy labels
|
🏛️ FDA Clearances K250288 - TeraRecon Cardiovascular.Calcification.CT: AI detecting and quantifying CT cardiovascular calcifications to enhance cardiac disease risk assessment. K250955 - XC11 ICE System: Intravascular ultrasound catheter for direct cardiac vascular imaging in structural heart and interventional procedures. K251827 - Azurion R3.1: Fluoroscopic X-ray system offering real-time interventional imaging in radiology-guided procedures. K251059 - Syngo Carbon Clinicals (Siemens): Automation software streamlining image processing and workflow for diagnostic radiologists. K252298 - ANDI 2.0: AI-based software for improved radiological image analysis and detection assistance. Explore last week's 8 radiology AI FDA approvals.
|
📄 Fresh Papers doi:10.1038/s41523-025-00831-x - Multi-timepoint mammography + risk factors model achieved 0.80 AUC for 10-year breast cancer risk on 9K+ women, outperforming single-exam and clinical models. doi:10.1016/j.jacr.2025.10.026 - GPT-4o outperformed GPT-4v and PGY-3 radiology residents on text-only board exam questions, but lagged behind trainees on image-based questions. doi:10.1007/s00261-025-05213-2 - Deep learning CT algorithm for osteoporosis (n=504 multicenter validation): AUC 0.96 internal, 0.82 external on noncontrast studies compared to DXA. doi:10.1186/s13244-025-02120-4 - 169 German referring physicians surveyed: 60% positive on AI in radiology, with transparency, legal clarity, and privacy top trust factors. Browse 205 new radiology AI studies from last week.
|
📰 Everything else in Radiology AI last week Researchers compared ChatGPT, Gemini, and Copilot LLMs for explaining radiology reports in lay terms for patients. MicroAdapt delivers 100,000x speed and 60% greater accuracy on small-footprint edge devices, enabling real-time adaptive AI for medical IoT. Mirai AI predicted up to 43% of interval breast cancers in 134,000+ UK patients from negative mammograms, enabling risk-adapted screening. Meta-analysis (6,600+ pts): AI tools surpassed radiologists in accuracy for post-treatment lung cancer imaging response assessment. MAGIC couples automated microscopy, AI imaging, and single-cell genomics to study cancer cell abnormalities in 100,000+ cells/day. Read 12 fresh radiology AI news articles for last week.
|
That's it for today! Before you go we’d love to know what you thought of today's newsletter to help us improve the RadAI Slice experience for you. |
|
👋 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. |
|
