TY - JOUR
T1 - Predictive Value of Deep Learning–derived CT Pectoralis Muscle and Adipose Measurements for Incident Heart Failure
T2 - Multi-Ethnic Study of Atherosclerosis
AU - Hathaway, Quincy
AU - Ibad, Hamza Ahmed
AU - Bluemke, David A.
AU - Pishgar, Farhad
AU - Kasaiean, Arta
AU - Klein, Joshua G.
AU - Cogswell, Rebecca
AU - Allison, Matthew
AU - Budoff, Matthew J.
AU - Barr, R. Graham
AU - Post, Wendy
AU - Bredella, Miriam A.
AU - Lima, João A.C.
AU - Demehri, Shadpour
N1 - Publisher Copyright:
© RSNA, 2023.
PY - 2023/10
Y1 - 2023/10
N2 - Purpose: To develop a deep learning algorithm capable of extracting pectoralis muscle and adipose measurements and to longitudinally investigate associations between these measurements and incident heart failure (HF) in participants from the Multi-Ethnic Study of Atherosclerosis (MESA). Materials and Methods: MESA is a prospective study of subclinical cardiovascular disease characteristics and risk factors for progression to clinically overt disease approved by institutional review boards of six participating centers (ClinicalTrials.gov identifier: NCT00005487). All participants with adequate imaging and clinical data from the fifth examination of MESA were included in this study. Hence, in this secondary analysis, manual segmentations of 600 chest CT examinations (between the years 2010 and 2012) were used to train and validate a convolutional neural network, which subsequently extracted pectoralis muscle and adipose (intermuscular adipose tissue (IMAT), perimuscular adipose tissue (PAT), extramyocellular lipids and subcutaneous adipose tissue) area measurements from 3031 CT examinations using individualized thresholds for adipose segmentation. Next, 1781 participants without baseline HF were longitudinally investigated for associations between baseline pectoralis muscle and adipose measurements and incident HF using crude and adjusted Cox proportional hazards models. The full models were adjusted for variables in categories of demographic (age, race, sex, income), clinical/laboratory (including physical activity, BMI, and smoking), CT (coronary artery calcium score), and cardiac MRI (left ventricular ejection fraction and mass (% of predicted)) data. Results: In 1781 participants (median age, 68 (IQR,61, 75) years; 907 [51%] females), 41 incident HF events occurred over a median 6.5-year follow-up. IMAT predicted incident HF in unadjusted (hazard ratio [HR]:1.14; 95%CI: 1.03–1.26) and fully adjusted (HR:1.16, 95%CI: 1.03–1.31) models. PAT also predicted incident HF in crude (HR:1.19; 95%CI: 1.06–1.35) and fully adjusted (HR:1.25; 95%CI: 1.07–1.46) models. Conclusion: The study demonstrates that fast and reliable deep learning-derived pectoralis muscle and adipose measurements are obtainable from conventional chest CT, which may be predictive of incident HF.
AB - Purpose: To develop a deep learning algorithm capable of extracting pectoralis muscle and adipose measurements and to longitudinally investigate associations between these measurements and incident heart failure (HF) in participants from the Multi-Ethnic Study of Atherosclerosis (MESA). Materials and Methods: MESA is a prospective study of subclinical cardiovascular disease characteristics and risk factors for progression to clinically overt disease approved by institutional review boards of six participating centers (ClinicalTrials.gov identifier: NCT00005487). All participants with adequate imaging and clinical data from the fifth examination of MESA were included in this study. Hence, in this secondary analysis, manual segmentations of 600 chest CT examinations (between the years 2010 and 2012) were used to train and validate a convolutional neural network, which subsequently extracted pectoralis muscle and adipose (intermuscular adipose tissue (IMAT), perimuscular adipose tissue (PAT), extramyocellular lipids and subcutaneous adipose tissue) area measurements from 3031 CT examinations using individualized thresholds for adipose segmentation. Next, 1781 participants without baseline HF were longitudinally investigated for associations between baseline pectoralis muscle and adipose measurements and incident HF using crude and adjusted Cox proportional hazards models. The full models were adjusted for variables in categories of demographic (age, race, sex, income), clinical/laboratory (including physical activity, BMI, and smoking), CT (coronary artery calcium score), and cardiac MRI (left ventricular ejection fraction and mass (% of predicted)) data. Results: In 1781 participants (median age, 68 (IQR,61, 75) years; 907 [51%] females), 41 incident HF events occurred over a median 6.5-year follow-up. IMAT predicted incident HF in unadjusted (hazard ratio [HR]:1.14; 95%CI: 1.03–1.26) and fully adjusted (HR:1.16, 95%CI: 1.03–1.31) models. PAT also predicted incident HF in crude (HR:1.19; 95%CI: 1.06–1.35) and fully adjusted (HR:1.25; 95%CI: 1.07–1.46) models. Conclusion: The study demonstrates that fast and reliable deep learning-derived pectoralis muscle and adipose measurements are obtainable from conventional chest CT, which may be predictive of incident HF.
UR - http://www.scopus.com/inward/record.url?scp=85175561293&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85175561293&partnerID=8YFLogxK
U2 - 10.1148/ryct.230146
DO - 10.1148/ryct.230146
M3 - Article
C2 - 37908549
AN - SCOPUS:85175561293
SN - 2638-6135
VL - 5
JO - Radiology: Cardiothoracic Imaging
JF - Radiology: Cardiothoracic Imaging
IS - 5
ER -