CT-derived pectoralis composition and incident pneumonia hospitalization using fully automated deep-learning algorithm: multi-ethnic study of atherosclerosis.

backgroundPneumonia-related hospitalization may be associated with advanced skeletal muscle loss due to aging (i.e., sarcopenia) or chronic illnesses (i.e., cachexia). Early detection of muscle loss may now be feasible using deep-learning algorithms applied on conventional chest CT.

objectivesTo implement a fully automated deep-learning algorithm for pectoralis muscle measures from conventional chest CT and investigate longitudinal associations between these measures and incident pneumonia hospitalization according to Chronic Obstructive Pulmonary Disease (COPD) status. MATERIALS AND

methodsThis analysis from the Multi-Ethnic Study of Atherosclerosis included participants with available chest CT examinations between 2010 and 2012. We implemented pectoralis muscle composition measures from a fully automated deep-learning algorithm (Mask R-CNN, built on the Faster Region Proposal Network (R-) Convolutional Neural Network (CNN) with an extension for mask identification) for two-dimensional segmentation. Associations between CT-derived measures and incident pneumonia hospitalizations were evaluated using Cox proportional hazards models adjusted for multiple confounders which include but are not limited to age, sex, race, smoking, BMI, physical activity, and forced-expiratory-volume-at-1 s-to-functional-vital-capacity ratio. Stratification analyses were conducted based on baseline COPD status.

resultsThis study included 2595 participants (51% female; median age: 68 (IQR: 61, 76)) CT examinations for whom we implemented deep learning-derived measures for longitudinal analyses. Eighty-six incident pneumonia hospitalizations occurred during a median 6.67-year follow-up. Overall, pectoralis muscle composition measures did not predict incident pneumonia. However, in fully-adjusted models, only among participants with COPD (N = 507), CT measures like extramyocellular fat index (hazard ratio: 1.98, 95% CI: 1.22, 3.21, p value: 0.02), were independently associated with incident pneumonia.

conclusionReliable deep learning-derived pectoralis muscle measures could predict incident pneumonia hospitalization only among participants with known COPD. CLINICAL RELEVANCE STATEMENT: Pectoralis muscle measures obtainable at zero additional cost or radiation exposure from any chest CT may have independent predictive value for clinical outcomes in chronic obstructive pulmonary disease patients. KEY POINTS: •Identification of independent and modifiable risk factors of pneumonia can have important clinical impact on patients with chronic obstructive pulmonary disease. •Opportunistic CT measures of adipose tissue within pectoralis muscles using deep-learning algorithms can be quickly obtainable at zero additional cost or radiation exposure. •Deep learning-derived pectoralis muscle measurements of intermuscular fat and its subcomponents are independently associated with subsequent incident pneumonia hospitalization.