Skip to main content
ACRADIO 2023 Annual Meeting Posters Main banner
Icon Legend
This session is not in your schedule.
This session is in your schedule. Click again to remove it.

Informatics and Data Science

Prediction of Heart Failure Incidence through Pectoralis Muscle Quality: Semantic Segmentation in Chest Computed Tomography

Saturday, May 6, 2023

    First Author(s)

  • QH

    Quincy A. Hathaway, PhD

    Medical Student
    West Virginia University School of Medicine, Heart and Vascular Institute
    Morgantown, West Virginia, United States

    • Co-Author(s)

  • HI

    Hamza A. Ibad, MD

    Postdoctoral Fellow
    Johns Hopkins University School of Medicine, Russell H. Morgan Department of Radiology and Radiological Science
    Baltimore, Maryland, United States

  • AK

    Arta Kasaiean, BS

    Undergraduate Research Assistant
    Johns Hopkins University School of Medicine, Russell H. Morgan Department of Radiology and Radiological Science
    Baltimore, Maryland, United States

  • JK

    Joshua Klein, BS

    Medical Student
    Johns Hopkins University School of Medicine, Russell H. Morgan Department of Radiology and Radiological Science
    Baltimore, Maryland, United States

  • RC

    Rebecca Cogswell, MD

    Associate Professor of Medicine
    University of Minnesota School of Medicine, Division of Cardiology
    Minneapolis, Minnesota, United States

  • JP

    James S. Pankow, PhD, MPH

    Professor, Division of Epidemiology and Community Health
    University of Minnesota School of Public Health, Division of Epidemiology and Community Health
    Minneapolis, Minnesota, United States

  • MB

    Matthew J. Budoff, MD

    Program Director and Director of Cardiac CT
    Harbor-UCLA Medical Center, Division of Cardiology
    Los Angeles, California, United States

  • DB

    David A. Bluemke, MD, PhD, MsB

    Professor, Thoracic Imaging and Cardiovascular Imaging Sections
    University of Wisconsin School of Medicine and Public Health, Department of Radiology
    Madison, Wisconsin, United States

  • JL

    João A. C. Lima, MD, MBA

    Professor, Medicine, Radiology and Epidemiology
    Johns Hopkins University School of Medicine, Division of Cardiology
    Baltimore, Maryland, United States

  • SD

    Shadpour Demehri, MD

    Professor of Radiology and Radiological Science
    Johns Hopkins University School of Medicine, Russell H. Morgan Department of Radiology and Radiological Science
    Baltimore, Maryland, United States

Purpose: Using computed tomography (CT), pectoralis muscle density and intramuscular fat deposition have been correlated with adverse health outcomes. It is unclear if pectoralis muscle measurements can be prospectively associated with incident heart failure and if the use of automated detection tools can bypass the need for cumbersome, manual measurements.

Methods/Materials: 3083 participants in the MESArthritis ancillary study were examined. Participants with absent baseline data (n=852), presence of heart failure at baseline (n=305), low-quality CT (n=26), or artifact (n=26) did not meet criteria, with 1874 participants included in the final analysis. Mediastinal windowing was applied and hamming distances, utilizing vantage point trees, selected the axial frame directly above the aorta. Manual measurements were performed on 600 participants using LabelMe. Patients were split into training (60%) and testing (40%) sets to develop our deep-learning model. The semantic segmentation platform was created with Pixellib and Mask R-CNN. Five pectoralis muscle composition indices were assessed: pectoralis muscle (PM), subcutaneous adipose tissue (SAT), intermuscular adipose tissue (IMAT), extramyocellular lipids (EML), and perimuscular adipose tissue (PAT).

Results: Comparison of manual measurements (n=600) to the deep-learning generated traces revealed a Dice score of (0.90 [0.90-0.91]) and (0.90 [0.89-0.90]) for the training and testing sets, respectively, and intersection over union score of (0.82 [0.82-0.83]) and (0.81 [0.81-0.82]) for the training and testing sets, respectively, when assessing all segments. There was no association between incident heart failure and PMAT (Hazard Ratio (HR): 0.96 [0.89-1.03]), EML (HR: 1.32 [0.82-2.11]), or SAT (HR: 0.99 [0.94-1.03]). Participants with higher IMAT (HR: 1.14 [1.01-1.29], P-value=0.047) and PAT (HR: 1.21 [1.03-1.43], P-value=0.021) had a higher risk of incident heart failure.

Conclusions: Measurements of inter- and intramuscular fat deposition can prospectively predict incident heart failure. The use of automated workflows can provide robust, reliable, and fast data in real-time at no additional cost or radiation exposure.