DeepACSA: Automatic Segmentation of Cross-Sectional Area in Ultrasound Images of Lower Limb Muscles Using Deep Learning

ABSTRACT Purpose Muscle anatomical cross-sectional area (ACSA) can be assessed using ultrasound and images are usually evaluated manually. Here, we present DeepACSA, a deep learning approach to automatically segment ACSA in panoramic ultrasound images of the human rectus femoris (RF), vastus lateralis (VL), gastrocnemius medialis (GM) and lateralis (GL) muscles. Methods We trained three muscle-specific convolutional neural networks (CNN) using 1772 ultrasound images from 153 participants (age = 38.2 yr, range = 13–78). Images were acquired in 10% increments from 30% to 70% of femur length for RF and VL and at 30% and 50% of muscle length for GM and GL. During training, CNN performance was evaluated using intersection-over-union scores. We compared the performance of DeepACSA to manual analysis and a semiautomated algorithm using an unseen test set. Results Comparing DeepACSA analysis of the RF to manual analysis with erroneous predictions removed (3.3%) resulted in intraclass correlation (ICC) of 0.989 (95% confidence interval = 0.983–0.992), mean difference of 0.20 cm 2 (0.10–0.30), and SEM of 0.33 cm 2 (0.26–0.41). For the VL, ICC was 0.97 (0.96–0.968), mean difference was 0.85 cm 2 (−0.4 to 1.31), and SEM was 0.92 cm 2 (0.73–1.09) after removal of erroneous predictions (7.7%). After removal of erroneous predictions (12.3%), GM/GL muscles demonstrated an ICC of 0.98 (0.96–0.99), a mean difference of 0.43 cm 2 (0.21–0.65), and an SEM of 0.41 cm 2 (0.29–0.51). Analysis duration was 4.0 ± 0.43 s (mean ± SD) for analysis of one image in our test set using DeepACSA. Conclusions DeepACSA provides fast and objective segmentation of lower limb panoramic ultrasound images comparable with manual segmentation. Inaccurate model predictions occurred predominantly on low-quality images, highlighting the importance of high-quality image for accurate prediction.