Issue #2
July 29, 2025

AI Models Diagnose Dementia Across Real-World Brain Images

PLUS: Study exposes demographic bias in leading skin disease diagnostic AI models

RadAI Slice Newsletter

Weekly Updates in Radiology AI

Good morning, there. A new AI model accurately distinguishes dementia types using over 300,000 real-world brain images.

Clinical AI models often falter when tested on large, diverse medical imaging data. This new approach demonstrates robust, generalizable performance across multiple sites, modalities, and patient populations. It marks vital progress toward trustworthy, practical neuroimaging AI for dementia diagnosis and care.


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

  • AI Diagnoses Dementia Types from Large, Heterogeneous Brain Imaging

  • Demographic Bias Persists in AI Skin Disease Diagnosis Tools

  • Open-Source AI Matches Commercial Tools for Radiology Reporting

  • Aidoc Raises $150M for Expansive Clinical AI Model Development

  • Plus: 6 FDA approved devices & 4 new papers.

LATEST DEVELOPMENTS

🧠 AI Diagnoses Dementia Types from Large, Heterogeneous Brain Imaging

RadAI Slice: A large-scale AI model can now differentiate major dementia types from diverse brain imaging datasets.

The details:

  • Trained on 308,000 brain scans from 17,000 patients across two decades

  • Detects Alzheimer's, Lewy body, vascular dementia, and more (AUC >0.84)

  • Handles MRI, CT, and PET images from varied real-world sources

  • Model structure accommodates different scan counts and types per patient

  • Validated across several hospital sites for operational robustness

Key takeaway: Real-world, multimodal imaging data have historically hindered AI clinical translation. This breakthrough shows scalable, generalizable imaging AI can perform robustly across different clinical settings, fostering wider adoption in dementia diagnostics.

👨‍⚕️ Demographic Bias Persists in AI Skin Disease Diagnosis Tools

RadAI Slice: ChatGPT-4 and LLaVA show varying fairness and accuracy diagnosing skin diseases, especially across age and sex groups.

The details:

  • Benchmarked AI models on 10,000 dermatoscopic skin disease images

  • ChatGPT-4 more fair for age and sex groups, LLaVA showed marked sex-bias

  • Both AI tools beat standard deep learning baselines in overall accuracy

  • Demographic fairness called essential before clinical AI deployment

Key takeaway: Despite rising accuracy, leading AI models can underperform for certain demographic groups. Proactive fairness evaluation is needed to prevent bias when deploying medical imaging AI in practice.

💡 Open-Source AI Matches Commercial Tools for Radiology Reporting

RadAI Slice: Hospitals can now deploy effective, privacy-conscious AI for radiology reporting using open-source models.

The details:

  • Synthetic radiology reports train models without real patient data

  • Open-source model Yi-34B matches GPT-4, smaller models pass GPT-3.5

  • Open-source deployment needs no data export, cutting privacy risks

  • Clinical workflow integration is practical and affordable

Key takeaway: Open-source AI enables cost-effective, scalable, and secure radiology report analysis—key for facilities concerned with data privacy and budgets.

💵 Aidoc Raises $150M for Expansive Clinical AI Model Development

RadAI Slice: Funding accelerates one of the most ambitious foundation models in radiology and healthcare AI.

The details:

  • Total Aidoc funding hits $370M, with $40M in credit

  • CARE targets 90% coverage of major diseases in three years

  • Backed by major hospitals, Nvidia, and AWS partnerships

  • Plans to support 45M+ patients annually across 150+ systems

Key takeaway: Massive funding propels foundation model development, with the potential to redefine AI's impact and scale in medical imaging and care delivery.

QUICK HITS

🏛️ FDA Clearances

  • K251151 - Rapid CTA 360 triages and notifies critical findings on CT angiography for urgent vascular care.

  • K250670 - EchoConfidence automates segmentation, function, and Doppler analysis for echocardiography exams.

  • K250484 - PIUR tUS inside adds 3D ultrasound volumes, segmentation, and characterization for thyroid imaging.

  • K243647 - Synapse PACS v7.5 features improved 3D visualization, AI-based bone removal, and semi-automatic segmentation.

  • K250999 - Samsung's V-series offers real-time diagnostic ultrasound for diverse patient care applications.

  • K251839 - uMI Panvivo PET/CT system provides advanced tomographic imaging for whole-body disease assessment.

📄 Fresh Papers

  • doi:10.1016/j.artmed.2025.103223 - Unsupervised machine learning on MRI tissue types predicts treatment response in liver disease trials.

  • arxiv:2507.16940v1 - AURA introduces an agentic AI that analyzes, explains, and evaluates medical images through dynamic interaction.

  • arxiv:2507.17662v1 - Mammo-Mamba blends state-space and transformer modules for efficient multi-view mammogram analysis.

  • doi:10.1016/j.jacadv.2025.102005 - AI-guided software enables novice nurses to acquire diagnostic-quality echocardiograms in real-world practice.

📰 Everything else in Radiology AI last week

That's it for today!

👋 We’d love your thoughts. Just hit Reply and tell us what you want to see in the next Slice!

See you next week.

Ready to Sharpen Your Edge?

Join hundreds of your peers who rely on RadAI Slice. Get the essential weekly briefing that empowers you to navigate the future of radiology.

We respect your privacy. Unsubscribe at any time.