AI flags early Alzheimer's on CT with 0.93 Dice accuracy

RadAISlice Newsletter

Weekly Updates in Radiology AI

Good morning, there. A deep learning CT model reached Dice >0.93 for brain atrophy across dementia stages.

I am struck by how CT-based segmentation could meaningfully expand dementia imaging where MRI remains inaccessible. This study establishes multicenter accuracy, setting a precedent for broader neurodegenerative disease monitoring. Such validated tools may foster earlier diagnosis and increase global equity in dementia care.

Would CT-based brain atrophy models change your dementia workup?

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

• CT AI Delivers MRI-Level Brain Atrophy Segmentation for Dementia

• Societies Issue Joint Guidelines for Safe Pediatric AI in Imaging

• AI Surpasses Expert Panel in Multicenter Pancreatic Cancer CT Detection

• Prospective, Multicenter Validation of MRI Deep Learning for TACE Response

• Plus: 1 newly released dataset & 4 new papers.

🧠 CT AI Delivers MRI-Level Brain Atrophy Segmentation for Dementia

RadAISlice: A validated nnU-Net model enables accurate dementia staging from routine head CT.

The details:

• 2337 CTs across Alzheimer's/FTD spectrum analyzed

• Dice coefficient exceeded 0.93 in key ventricular regions

• Correlations to MRI volumetry up to 0.996

• External validation revealed stage-specific atrophy patterns

Key takeaway: CT-based neurodegeneration metrics could broaden access to dementia diagnostics, offering interpretable, robust segmentation where MRI is unavailable.

🛡️ Societies Issue Joint Guidelines for Safe Pediatric AI in Imaging

Image from: Radiology Business

RadAISlice: Leading societies collaborate to publish child-centered AI guidance for imaging.

The details:

• Six pediatric/adult radiology groups signed the policy

• Warns most FDA-cleared AI is adult-focused

• Calls for regulation, outcome tracking, and pediatric validation

• Addresses purchasing, integration, interpretation, and education steps

Key takeaway: Policy now pushes for pediatric-specific validation and outcome monitoring before AI adoption, impacting developers and imaging departments globally.

🔬 AI Surpasses Expert Panel in Multicenter Pancreatic Cancer CT Detection

RadAISlice: Paired-reader study shows AI outperforming radiologists for PDAC detection.

The details:

• 1130 patients, robust gold standards, external test sites

• AI AUC 0.92 vs. radiologist pool 0.88 (p=0.001)

• Superior results for both tumor detection and diagnostic speed

• Open-source benchmark released for future validation

Key takeaway: AI can meaningfully raise detection rates for late-presenting cancers like PDAC, supporting adoption in high-stakes diagnostic settings.

🩺 Prospective, Multicenter Validation of MRI Deep Learning for TACE Response

RadAISlice: DL model using pretreatment MRI predicts post-TACE response and prognosis.

The details:

• Multicenter: data from 3 hospitals, hundreds of HCC patients

• DL-based MRI and MLP combine for AUC 0.80–0.82

• Performance confirmed on two external test sets

• DL features linked with angiogenesis/hypoxia gene signatures

Key takeaway: Prospective external validation supports integrating MRI-based DL for early, noninvasive HCC therapy response prediction across diverse practice settings.

SonoUS Public Datasets Collection (2025-09)

Modality: US | Focus: Multiple, including heart and brain | Task: Segmentation, classification

• Size: 72 datasets, ranging from ~5 to 83,000+ scans; 10 to 75,000+ patients/entities per set

• Annotations: Various; includes organ/lesion labels, segmentations, and structured metadata

• Institutions: University of Oxford, Khalifa University et al.

• Availability:

  Public (webpage link forthcoming)

• Highlight: Largest public catalog of ultrasound datasets, scored for openness (SonoDQS) and model transparency (SonoMQS); covers wide anatomical and worldwide data diversity.

📄 Fresh Papers

• doi:10.1136/bmjresp-2025-003510 - Two-center, externally validated CT study: AI-based pulmonary vessel features predict rapid ILD progression and prognosis in myositis patients.

• doi:10.1016/j.rmed.2025.108545 - COPDGene study: CT-based AI detects 85% of under-diagnosed interstitial lung disease at 98% specificity in a population cohort.

• doi:10.1007/s10140-025-02420-8 - Thematic review explores emergency radiology's transformation by AI, workforce shortages, and workflow integration challenges.

• doi:10.1186/s12931-025-03411-6 - Prospective single-center study: ML model integrating bedside multimodal ultrasound achieves AUC 0.93 for COPD exacerbation diagnosis.

• Browse 193 new radiology AI studies from last week.

📰 Everything else in Radiology AI last week

• Major AI investments and expansion from Nvidia, AWS and Canon highlight advances—and persistent challenges—in radiology, genomics, and clinical operations.

• Nvidia CEO states AI is enhancing, not replacing, radiology careers as clinical demand grows.

• Nvidia and Amazon's $50B AI push in genomics and radiology spotlights the growing urgency for radiology professionals to adapt.

• A CT-based AI model quantifies adrenal gland volume to correlate with chronic stress levels, validated by biomarkers and questionnaires.

• Image-only mammography AI model (421,500 scans, 27 centers) predicts 5-year cancer risk more accurately than breast density.

• Read 11 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.