🚨 New work in AI-driven hypertrophic cardiomyopathy (HCM) management
Beyond Doppler: scalable AI detection of LVOT obstruction in HCM
Collaboration between Mayo Clinic Cardiology and AI innovation hub
We developed a multi-view deep learning model to detect LVOT obstruction (>20 mmHg) using routine 2D echocardiography—no Doppler required.
Key idea: exploit spatial-temporal signals across standard echo views.
📊 Results:
• Internal cohort (N=1833), external validation (Korea, N=46)
• Single-view models limited (AUC 0.47–0.70)
• Echo foundation model (EchoPrime) improved performance (AUC up to 0.83)
• Multi-view fusion further boosted performance to AUC 0.84 externally
💡 Takeaway: LVOT physiology is embedded in routine 2D imaging and can be reliably extracted via cross-view AI fusion.
Potential impact: reduced dependence on Doppler expertise, scalable LVOT assessment, and real-time decision support for broader and resource-limited settings. Chadi Ayoub Reza Arsanjani Juan Maria Farina MD Owen Crystal, M.S.
https://lnkd.in/gKXT3r5D
Recent news
📢 New Editorial for our paper: Turning our Incidental BAC detection into a Cardiovascular Prevention Opportunity https://lnkd.in/dijir3HX
They discuss how breast arterial calcification (BAC) on mammograms can serve as an opportunistic marker of vascular aging and cardiometabolic risk.
Propose a simple two-lane prevention framework:
• ASCVD prevention: optimize lipids, BP, diabetes, lifestyle
• Heart failure prevention: recognize BAC as a signal of vascular stiffness and HF risk
The key step is simple: report BAC consistently and use it to trigger cardiovascular risk assessment.
Proud of our student Ramon Correa and Jiwoong Jeong for completing PhD Arizona State University
🚨 New Paper in hashtag#EuropeanRadiology | Robust Kidney Segmentation Across hashtag#CT Domains
AI models often fail where it matters most — real-world variability. In kidney CT imaging, even simple shifts like contrast phase changes can significantly degrade segmentation performance.
🔬 In our latest work published in European Radiology, we tackle this challenge head-on.
👉 We introduce a domain adaptation framework powered by a latent space discriminator that enables robust kidney segmentation across:
Contrast vs. non-contrast scans
Arterial vs. venous phases
Normal vs. abnormal kidneys
📊 Key results:
✅ Validated across KiTS21, STU, and Mayo Clinic datasets
✅ Achieved DICE = 0.8892 under domain shift
✅ Outperformed strong baselines including hashtag#TotalSegmentator
✅ Required less training data while improving generalization
💡 Why this matters:
Reliable kidney segmentation shouldn’t depend on imaging protocol variations. Our approach enables consistent volumetric assessment, supporting better diagnosis and personalized care — regardless of contrast differences or pathology.
🔓 We’re also committed to open science — code is publicly available.
💬 Would love to hear thoughts from the community on domain generalization in medical imaging — how are you addressing distribution shifts in your work?
hashtag#AI hashtag#Radiology hashtag#MedicalImaging hashtag#DeepLearning hashtag#DomainAdaptation hashtag#KidneyDisease hashtag#HealthcareAI hashtag#OpenScience https://lnkd.in/g3MzD6Yt
Our#ACC26 poster on multimodal AI for cardiovascular risk prediction 🚀
We present a parameter-efficient framework for 5-year major adverse cardiovascular event (#MACE) prediction that fuses opportunistic non-gated #CT imaging (contrast + non-contrast) with free-form clinical narratives using cross-modal attention. A practical path toward real-world deployment for early identification of high-risk patients. https://www.jacc.org/doi/abs/10.1016/j.jacc.2026.02.3014
https://www.linkedin.com/posts/imon-banerjee-017060a0_open-source-hybrid-large-language-model-integrated-activity-7345959229205159936-NEmj?utm_source=share&utm_medium=member_desktop&rcm=ACoAABVdZhMBKMI0vbwj-o-lI8L4ldXRTYbha6I