Deep Learning Technique for Musculoskeletal Analysis

Abstract

Advancements in musculoskeletal analysis have been achieved by adopting deep learning technology in image recognition and analysis. Unlike musculoskeletal modeling based on computational anatomy, deep learning-based methods can obtain muscle information automatically. Through analysis of image features, both approaches can obtain muscle characteristics such as shape, volume, and area, and derive additional information by analyzing other image textures. In this chapter, we first discuss the necessity of musculoskeletal analysis and the required image processing technology. Then, the limitations of skeletal muscle recognition based on conventional handcrafted features are discussed, and developments in skeletal muscle recognition using machine learning and deep learning technology are described. Next, a technique for analyzing musculoskeletal systems using whole-body computed tomography (CT) images is shown. This study aims to achieve automatic recognition of skeletal muscles throughout the body and automatic classification of atrophic muscular disease using only image features, to demonstrate an application of whole-body musculoskeletal analysis driven by deep learning. Finally, we discuss future development of musculoskeletal analysis that effectively combines deep learning with handcrafted feature-based modeling techniques.

Keywords: 2D U-Net; 3D U-Net; Deep muscle; FCN-8s; Musculoskeletal analysis; Musculoskeletal segmentation; Random forest; Skeletal muscle; Surface muscle.