Development of Fully Automated 3D Pectoralis Muscle Measurements on Computed Tomography Imaging for Chronic Obstructive Pulmonary Disease Using Deep Learning

Abstract Rationale: The pectoralis is commonly used as a surrogate for measuring muscle composition in computed tomography (CT) images in chronic obstructive pulmonary disease (COPD). Two-dimensional (2D) pectoralis muscle area (PMA) measurements at the aortic arch slice are associated with lung function and exercise capacity in COPD. However, single-slice measurements are prone to variations in anatomy and may not be representative of the 3D structure. Additionally, a highly accurate and externally validated, automated 3D pectoralis muscle volume (PMV) segmentation pipeline has yet to be developed. The objective of this study was to develop a deep-learning model to segment the PMV from chest CT images and compare the association between PMA and PMV measurements with lung function and exercise capacity in those with COPD. Methods: Participants with/without spirometrically-defined COPD and CT images from the multicenter Canadian Cohort of Obstructive Lung Disease study (NCT00920348) were included. A subset of participants were randomly sampled, with 8/9 sites acting as the development cohort (N=96 training; N=16 validation), and the last site serving as the internal testing cohort (N=32). COPD patients from the University of Alberta (NCT03679312) served as an external testing cohort (N=32). Ground-truth segmentations were manually segmented by an experienced image-analyst, and a 3D U-Net model was trained. Model performance was determined using the Dice-Sorensen Coefficient (DSC). For PMA measurements, the aortic arch was automatically detected using TotalSegmentator (10.1148/ryai.230024). Differences between the PMA (cm2) and PMV (cm3) measurements of no-COPD/COPD participants and associations with forced expiratory volume in 1-second (FEV1 (L)) and peak rate of oxygen consumption (V'O2peak (mL/kg/min)) during cardiopulmonary exercise testing were quantified using an ANCOVA and linear regression models adjusted for age, sex, height, BMI, center ID, and comorbidities. Results: 1325 participants had automated PMV segmentations (no-COPD (N=684): age=65.8±9.6yrs, 311(45.5%) females, FEV1=101±17%pred; COPD (N=641): age=67.3±10.1yrs, 243(37.9%) females, FEV1=82±19%pred). The trained model obtained a DSC of 0.94±0.02, 0.92±0.03, 0.92±0.03, and 0.92±0.04 on the training, validation, internal and external testing datasets, respectively. Compared to no-COPD participants, COPD participants had smaller PMA (34.7±11.6 vs. 37.0±10.7cm2; p<0.05) and PMV measurements (488.9±145.7 vs. 505.2±133.1cm3; p<0.05). PMV measurements showed stronger associations than PMA with FEV1 (Radj2=0.58 vs. 0.56, p<0.05) and V'O2peak (Radj2=0.57 vs. 0.55, p<0.05). Conclusions: PMV can be automatically extracted from CT images with high accuracy, quantify COPD-associated muscle atrophy, and outperforms PMA for associations with airflow obstruction and exercise intolerance. These findings support the use of PMV as a biomarker of obstructive lung-disease related muscle-loss.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.