About

Michael Salerno is a professor of medicine in the field of cardiovascular medicine and also holds a position in radiology specializing in cardiovascular imaging at Stanford University. He is affiliated with the Center for Artificial Intelligence in Medicine & Imaging (AIMI), where his work focuses on integrating artificial intelligence into medical and imaging research to advance healthcare. His contributions include applying AI techniques to improve cardiovascular diagnostics and imaging, leveraging interdisciplinary expertise to develop innovative solutions in medical imaging and cardiovascular medicine.

Research topics

• Internal medicine
• Medicine
• Cardiology
• Endocrinology
• Genetics
• Radiology
• Biology
• Surgery

Selected publications

• 3D spiral-navigated simultaneous Cine, LGE and T1 mapping under free-breathing and self-gating (CAT-SPARCS) using dual-excitation flip-angle with randomized stack-of-spiral

  Journal of Cardiovascular Magnetic Resonance · 2026-01-01

  articleOpen accessSenior author
  PublisherDOI

• Motion-Augmented End-to-End Nonrigid MOCO-DESIRE for Free-Breathing Spiral First-Pass Myocardial Perfusion Imaging at 3T with Limited Training Data

  Journal of Cardiovascular Magnetic Resonance · 2026-01-01

  articleOpen accessSenior author
  PublisherDOI

• MARCUS: An agentic, multimodal vision-language model for cardiac diagnosis and management

  ArXiv.org · 2026-03-23

  articleOpen access

  Cardiovascular disease remains the leading cause of global mortality, with progress hindered by human interpretation of complex cardiac tests. Current AI vision-language models are limited to single-modality inputs and are non-interactive. We present MARCUS (Multimodal Autonomous Reasoning and Chat for Ultrasound and Signals), an agentic vision-language system for end-to-end interpretation of electrocardiograms (ECGs), echocardiograms, and cardiac magnetic resonance imaging (CMR) independently and as multimodal input. MARCUS employs a hierarchical agentic architecture comprising modality-specific vision-language expert models, each integrating domain-trained visual encoders with multi-stage language model optimization, coordinated by a multimodal orchestrator. Trained on 13.5 million images (0.25M ECGs, 1.3M echocardiogram images, 12M CMR images) and our novel expert-curated dataset spanning 1.6 million questions, MARCUS achieves state-of-the-art performance surpassing frontier models (GPT-5 Thinking, Gemini 2.5 Pro Deep Think). Across internal (Stanford) and external (UCSF) test cohorts, MARCUS achieves accuracies of 87-91% for ECG, 67-86% for echocardiography, and 85-88% for CMR, outperforming frontier models by 34-45% (P<0.001). On multimodal cases, MARCUS achieved 70% accuracy, nearly triple that of frontier models (22-28%), with 1.7-3.0x higher free-text quality scores. Our agentic architecture also confers resistance to mirage reasoning, whereby vision-language models derive reasoning from unintended textual signals or hallucinated visual content. MARCUS demonstrates that domain-specific visual encoders with an agentic orchestrator enable multimodal cardiac interpretation. We release our models, code, and benchmark open-source.

  PublisherOA PDF

• Cardiac Strain Imaging in Asymptomatic Carriers of Transthyretin Variants

  Research Square · 2026-02-12

  preprintOpen access
  PublisherOA PDFDOI

• Quantitative Cardiac Perfusion with Kolmogorov--Arnold Networks for Robust Myocardial Blood Flow Mapping

  Journal of Cardiovascular Magnetic Resonance · 2026-01-01

  articleOpen accessSenior author
  PublisherDOI

• Biobank-Scale Plasma Proteomics Identifies Novel Biomarkers in Hypertrophic Cardiomyopathy

  Circulation Genomic and Precision Medicine · 2026-04-22

  articleOpen access

  BACKGROUND: Hypertrophic cardiomyopathy (HCM) is characterized by substantial heterogeneity in both clinical phenotype and risk of adverse outcomes, including heart failure and sudden cardiac death. This highlights the need for robust biomarkers for risk stratification, and while previous studies have identified the role of select plasma proteins, comprehensive large-scale proteomic analyses have been limited in HCM. METHODS: We performed a case-control analysis of 2922 plasma proteins in 49 588 UK Biobank participants (100 HCM cases) to identify proteins associated with HCM. External replication analyses were performed in the deCODE Genetics Icelandic study (51 cases/38 904 controls) and All of Us (546 cases/41 049 controls) data sets. Associations with adverse clinical outcomes and cardiac endophenotypes of disease severity were further identified, and causal relationships were evaluated using Mendelian randomization. Relative biomarker importance was also assessed by joint modeling via machine learning. RESULTS: We identified novel associations of ANGPT2 (angiopoietin-2) and LTBP2 (latent transforming growth factor-beta binding protein 2) with HCM, with both also showing prognostic utility for heart failure-related outcomes in HCM cases. We also confirmed the associations of established biomarkers (eg, NT-proBNP [N-terminal pro-B-type natriuretic peptide], troponins I and T) with HCM cases, cardiac imaging markers of disease severity, and adverse outcomes. Mendelian randomization analyses supported a causal effect of HCM on increasing NT-proBNP and troponin T levels. CONCLUSIONS: This biobank-scale plasma proteomic study in HCM identified ANGPT2 and LTBP2 as novel HCM biomarkers with potential diagnostic and prognostic utility. These findings highlight the potential for plasma proteomics to improve risk prediction and provide insight into HCM pathobiology.

  PublisherDOI

• Plug-and-play with Denoising Diffusion Probabilistic Model (DDPM) Denoiser for Highly Accelerated Cardiac Perfusion Image Reconstruction

  Journal of Cardiovascular Magnetic Resonance · 2026-01-01

  articleOpen accessSenior author
  PublisherDOI

• Optimizing breast and chest wall treatment planning: Integrating dynamic collimator rotation with static‐angle modulated ports in VMAT radiotherapy

  Journal of Applied Clinical Medical Physics · 2026-02-01

  articleOpen access

  PURPOSE: RapidArc Dynamic (RAD) integrates static-angle modulated ports (STAMPs) and a dynamic collimator into arc delivery. The optimal use of RAD, including the ideal number of STAMPs, the best use of the dynamic collimator, and the ideal relative weighting between arc and STAMPs, has yet to be reported. We aim to investigate optimized utility of these parameters for breast and chest wall treatment planning to achieve superior dosimetric results. METHODS: Thirteen breast and chest wall patients were planned using RAD. Plans were created using the three different dynamic collimator options, five different arc/STAMP weighting options, and with 2, 4, and 6 STAMPs. All plans were created with automated skin flash. RAD plans were compared to conventional RapidArc (RA) plans. The DVH metrics and MUs for each plan were recorded, and a paired T-test was used to test for statistically significant (p ≤ 0.05) differences between the plans. RESULTS: "Optimize between static angles" was the best option for dynamic collimator setting. Increasing the number of STAMPs from 2 to 4 or 6 lowered PTV V105% in patients where the PTV V105% was high but provided limited benefit in most patients. Selecting arc-dominant weighting yields significantly worse DVH metrics than a balanced weighting. Dosimetric differences were minimal between (0) Balanced, (1) Static, or (2) Static-Dominant weighting. CONCLUSIONS: The following are recommended as a starting point for breast and chest wall RAD plans: 2 STAMPs positioned similar to breast tangents, "optimize between static angles" for the dynamic collimator, and a weighting of either (0) balanced, (1) static, or (2) static-dominant. The arc-dominant setting resulted in plans of the lowest quality.

  PublisherOA PDFDOI

• T1 Troubles: Ironing Out Abnormal T1 Values in a Case of Infiltrative Cardiomyopathy

  Journal of Cardiovascular Magnetic Resonance · 2026-01-01

  articleOpen accessSenior author
  PublisherDOI

• Multimodal Machine Learning Reveals the Genomic and Proteomic Architecture of Heart Failure with Preserved Ejection Fraction

  medRxiv · 2026-02-09

  articleOpen access

  Abstract Heart failure with preserved ejection fraction (HFpEF) affects over 30 million people and lacks disease-modifying therapies. Although genomic-led drug discovery increases success by more than 2.6-fold, HFpEF genomic discovery remains constrained by imprecise phenotyping in biobanks, with only two loci identified to date. Biobanks lack HFpEF diagnostic codes and echocardiograms, yet HFpEF diagnosis exists along a continuum and is inherently probabilistic, presenting an opportunity for multimodal prediction. Here we introduce TRI-modal Assessment and Discovery of HFpEF (TRIAD-HFpEF), a machine learning framework integrating electrocardiograms, cardiac magnetic resonance imaging, and biomarkers to assign HFpEF probabilities. Deployed in UK Biobank, these probabilities validate with respect to mortality, hospitalizations, and structural and functional HFpEF features. Genome-wide and proteomic analyses reveal over 90 novel loci, a 45-fold expansion, and distinguish causal proteins from non-causal biomarkers of disease progression, prioritizing 11 therapeutic targets and 7 non-causal biomarkers. We identify FLT3 as one of the 11 therapeutic targets, consistent with the reported 7-fold increased heart failure risk from FLT3 inhibitors in leukemia. We validate this finding by demonstrating significant worsening of diastolic function following FLT3 inhibitor treatment in an independent clinical cohort. Conversely, MPO emerged as one of the 7 non-causal biomarkers, aligning with three recent negative MPO inhibitor trials. TRIAD-HFpEF demonstrates that machine learning-derived phenotypes can unlock genetic discovery in complex syndromes, identifying actionable targets while deprioritizing associations reflecting disease consequences rather than causes.

  PublisherOA PDFDOI

Recent grants

• Rapid Free-Breathing Self-Gated Spiral Pulse Sequences for Simultaneous Cine and T1 mapping

  NIH · $558k · 2021–2021

• High-Resolution Whole Heart Quantitative CMR Perfusion Imaging in Ischemic Heart Disease

  NIH · $2.9M · 2017–2025

• NIH Grant K23HL11291001

  NIH · $136k · 2017

• Rapid Free-Breathing Self-Gated Spiral Pulse Sequences for Simultaneous Cine and T1 mapping

  NIH · $1.6M · 2021–2027

• NIH Grant K23HL112910

  NIH · $653k · 2017

Frequent coauthors

• Christopher M. Kramer

  Sequoia (United States)

  179 shared

• Frederick H. Epstein

  77 shared

• Craig H. Meyer

  64 shared

• Stefan Neubauer

  John Radcliffe Hospital

  64 shared

• Yang Yang

  Nantong University

  56 shared

• Debiao Li

  55 shared

• Håkan Arheden

  Lund University

  51 shared

• Andreas Kumar

  Health Sciences North

  49 shared