Deep Learning-Based 3D and 2D Approaches for Skeletal Muscle Segmentation on CT Images

Sarcopenia is a clinical condition characterized by the progressive loss of skeletal muscle mass and strength. It can be quantitatively assessed through the analysis of Computed Tomography (CT) images. Accurate segmentation of skeletal muscle from CT scans is crucial for ensuring an objective, standardized, and reproducible diagnosis. Although manual segmentation is currently applied as the reference standard, it is inherently time-consuming and not feasible for large-scale or routine clinical application. In this study, we propose two deep learning (DL) models for the automatic segmentation of skeletal muscle at the level of the third lumbar vertebra (L3): a three-dimensional model based on UNet3+ and a two-dimensional one based on DeepLabv3+. Both models have been trained and evaluated on a dataset of 430 patients, using a combination of qualitative and quantitative metrics, such as Dice Similarity Coefficient (DSC), 95th percentile Hausdorff Distance (HD95), Average Surface Distance (ASD), sensitivity, and specificity. The DeepLabv3+ model achieved an average DSC of 0.982 and an HD95 of 1.04 mm, while the UNet3+ model reached a DSC of 0.967 and an HD95 of 1.27 mm, despite employing a significantly lower number of parameters. The two proposed models demonstrated high accuracy in muscle segmentation. The 3D UNet3+ model combines strong performance with low computational impact, proving the potential clinical applicability of volumetric segmentation approaches in the quantitative assessment of muscle mass.