An Analysis of the Efficacy of Deep Learning-Based Pectoralis Muscle Segmentation in Chest CT for Sarcopenia Diagnosis.

Abstract

Sarcopenia is the loss of skeletal muscle function and mass and is a poor prognostic factor. This condition is typically diagnosed by measuring skeletal muscle mass at the L3 level. Chest computed tomography (CT) scans do not include the L3 level. We aimed to determine if these scans can be used to diagnose sarcopenia and thus guide patient management and treatment decisions. This study compared the ResNet-UNet, Recurrent Residual UNet, and UNet3 + models for segmenting and measuring the pectoralis muscle area in chest CT images. A total of 4932 chest CT images were collected from 1644 patients, and additional abdominal CT data were collected from 294 patients. The performance of the models was evaluated using the dice similarity coefficient (DSC), accuracy, sensitivity, and specificity. Furthermore, the correlation between the segmented pectoralis and L3 muscle areas was compared using linear regression analysis. All three models demonstrated a high segmentation performance, with the UNet3 + model achieving the best performance (DSC 0.95 ± 0.03). Pearson correlation coefficient between the pectoralis and L3 muscle areas showed a significant positive correlation (r = 0.65). The correlation coefficient between the transformed pectoralis and L3 muscle areas showed a stronger positive correlation in both univariate analysis using only muscle area (r = 0.74) and multivariate analysis considering sex, weight, age, and muscle area (r = 0.83). Segmentation of the pectoralis muscle area using artificial intelligence (AI) on chest CT was highly accurate, and the measured values showed a strong correlation with the L3 muscle area. Chest CT using AI technology could play a significant role in the diagnosis of sarcopenia.