Automated CT-derived skeletal muscle mass determination in lower hind limbs of mice using a 3D U-Net deep learning network

Abstract

The loss of skeletal muscle mass is recognized as a complication of several chronic diseases and is associated with increased mortality and a decreased quality of life. Relevant and reliable animal models in which muscle wasting can be monitored noninvasively over time are instrumental to investigate and develop new therapies. In this work, we developed a fully automatic deep learning algorithm for segmentation of micro cone beam computed tomography images of the lower limb muscle complex in mice and subsequent muscle mass calculation. A deep learning algorithm was trained on manually segmented data from 32 mice. Muscle wet mass measurements were obtained from 47 mice and served as a data set for model validation and reverse model validation. The automatic algorithm performance was ~150 times faster than manual segmentation. Reverse validation of the algorithm showed high quantitative metrics (i.e., a Dice similarity coefficient of 0.93, a Hausdorff distance of 0.4 mm, and a center of mass displacement of 0.1 mm), substantiating the robustness and accuracy of the model. A high correlation ( R2 = 0.92) was obtained between the computed tomography-derived muscle mass measurements and the muscle wet masses. Longitudinal follow-up revealed time-dependent changes in muscle mass that separated control from lung tumor-bearing mice, which was confirmed as cachexia. In conclusion, this deep learning model for automated assessment of the lower limb muscle complex provides highly accurate noninvasive longitudinal evaluation of skeletal muscle mass. Furthermore, it facilitates the workflow and increases the amount of data derived from mouse studies while reducing the animal numbers. NEW & NOTEWORTHY This deep learning application enables highly accurate noninvasive longitudinal evaluation of skeletal muscle mass changes in mice with minimal requirement for operator involvement in the data analysis. It provides a unique opportunity to increase and analyze the amount of data derived from animal studies automatically while reducing animal numbers and analytical workload.