C108-20 Automated Deep Learning Quantification of Pectoralis Muscle Area in COPDGene: Validation Against Manual Reference Standards

Abstract Introduction Loss of skeletal muscle mass, an important systemic manifestation of COPD, is linked to functional decline and mortality. Pectoralis muscle area (PMA) measured from chest CT is a reproducible biomarker of body composition, but manual annotation remains time-consuming and prone to inter-reader variability. We developed and validated a convolutional neural network (CNN) for automated quantification of pectoralis major and minor muscle areas from chest CT scans and compared AI-derived values to expert manual measurements. Methods Development: To generate training and validation data, the entire TotalSegmentatorv2 dataset (training) and n = 150 COPDGene scans (validation) were processed by TotalSegmentator segmentation workflows. The model architecture consisted of a 3D CNN trained on CT volumes cropped to the lungs and downsampled to 1.5mm3. Training used patchshuffle regularization and the Adam optimizer (learning rate 0.0001, Dice focal loss). PMA validation dice coefficient was 0.85 on n = 150 COPDGene scans. Performance and associations with clinical data: We used a total of n = 8732 COPDGene Phase 1 inspiratory CT scans. Performance was assessed by correlation and Bland-Altman analysis against previously generated manual PMA measurements, repeating analyses before and after calibration (bias correction using linear regression). We evaluated associations between PMA and key COPD-related outcomes using multivariable linear or logistic regression, including FEV1% predicted, FVC% predicted, FEV1/FVC ratio, percent emphysema, St. George’s Respiratory Questionnaire (SGRQ) scores, resting oxygen saturation, six-minute walk distance, BODE index, modified MRC dyspnea score, and exacerbations. Manual and AI-derived PMA were independent variables adjusted for age at enrollment, sex, height, smoking status, and smoking pack-years. We calculated baseline-only and time-varying models (incorporating PMA from all-time points), unadjusted and adjusted for age at enrollment, sex, BMI, smoking status, and GOLD stage. To determine the association between PMA (standardized per 1 SD) and all-cause mortality, we employed Cox proportional hazards models. Results AI-derived PMA strongly correlated with manual measurements Pearson correlation = 0.934 (p 0.001); bias±LOA 7.4±12.5cm² (Fig. 1a). After calibration, systematic bias was nearly eliminated, with a bias±LOA 0.4±11.4cm² (Fig. 1b). Automated and manual measurements had similar associations with COPD-related outcomes when evaluated using multivariable linear or logistic regression (Table 1). Smaller PMA was (automated method) significantly associated with mortality (Fig. 1c). Conclusions Automated PMA measurements accurately reproduce manual PMA measurements and enables high-throughput assessment of chest CT body composition. Future work will extend the approach to longitudinal PMA change and explore muscle density as a marker of systemic disease severity. This abstract is funded by: This work was supported by NHLBI grants U01 HL089897 and U01 HL089856 and by NIH contract 75N92023D00011