Ultrasound Segmentation Using Deep Learning: Training on Musculoskeletal Phantom Data and Testing on Clinical Data

Muscle segmentation in musculoskeletal (MSK) ultrasound (US) images is a crucial task for various medical applications. However, manual segmentation can be time-consuming and prone to errors. Deep learning models are promising in automating this task; yet, developing effective models is difficult when the anatomical regions of interest have limited data for training. We addressed this challenge by exploring the potential of training deep learning models using MSK US phantom images and applying these models to clinical images for muscle segmentation. We conducted a thorough performance comparison of various deep learning models on both phantom data and clinical data. Our study showcased the transferability of phantom-trained model to clinical data, highlighting the potential of using phantom data as an alternative training dataset when clinical data is scarce. This study established a promising approach for efficient and accurate muscle segmentation in clinical scenarios.