Automatic anatomical cross-sectional area measurements in human lower limb muscle ultrasound images using deep learning

Background: Muscle anatomical cross-sectional area (ACSA) is an important parameter characterizing muscle function and the severity of several muscular disorders. Ultrasonography is a patient friendly, fast, and cheap method of assessing muscle ACSA, however manual image analysis is laborious, subjective and requires thorough experience. Convolutional neural networks (CNNs) demonstrated promising results for muscle ultrasonography analysis. To date, no open-access and fully automated program to segment ACSA in panoramic ultrasound images is available.
Methods: In this thesis, DeepACSA, a deep learning approach to automatically segment ACSA of the human rectus femoris (RF), vastus lateralis (VL), gastrocnemius medialis (GM) and lateralis (GL) muscles is presented and discussed. We trained three muscle-specific CNNs using 1772 images from 153 participants (25 females, 128 males; mean age: 38.2 years, range: 13-78) captured by three experienced operators using three distinct devices.
Results: Comparing DeepACSA analysis of the RF to manual analysis resulted in intra-class correlation (ICC) of 0.96 (95% CI 0.94,0.97), mean difference of 0.31 cm2 (0.04,0.58) and standard error of the differences (SEM) of 0.91 cm2 (0.47,1.36). For the VL, ICC was 0.94
(0.91,0.96), mean difference was 0.25 cm2 (-0.21,0.7) and SEM was 1.55 cm2 (1.13,1.96). The GM/GL muscles demonstrated an ICC of 0.97 (0.95,0.98), a mean difference of 0.01 cm2 (-0.25,0.24) and a SEM of 0.69 cm2 (0.52,0.83).
Conclusion: DeepACSA provides segmentation of lower limb muscle ACSA in panoramic ultrasound images comparable to manual segmentation. Yet, future versions of DeepACSA should include CNN performance comparisons, clinical populations and increased user-friendliness.