Automatic Tracking of Muscle Cross‐Sectional Area Using Convolutional Neural Networks with Ultrasound

OBJECTIVES: The purpose of this study was to develop an automatic tracking method for the muscle cross-sectional area (CSA) on ultrasound (US) images using a convolutional neural network (CNN). The performance of the proposed method was evaluated and compared with that of the state-of-the art muscle segmentation method. METHODS: A real-time US image sequence was obtained from the rectus femoris muscle during voluntary contraction. A CNN was built to segment the rectus femoris muscle and calculate the CSA in each US frame. This network consisted of 2 stages: feature extraction and score map reconstruction. The training of the network was divided into 3 steps with output score map resolutions of one-fourth, one-half, and all of the original image. We evaluated the segmentation performance of our method with 5-fold cross-validation. The mean precision, recall, and dice similarity score were calculated. RESULTS: The mean precision, recall, and Dice's coefficient (DSC) ± SD were 0.936 ± 0.029, 0.882 ± 0.045, and 0.907 ± 0.023, respectively. Compared with the state-of-the-art muscle segmentation method (constrained mutual-information-based free-form deformation), the proposed method using CNN showed high performance. CONCLUSIONS: The automated method proposed in this study provides an accurate and efficient approach to the estimation of the muscle CSA during muscle contraction.