Building a machine learning-based platform to analyze skeletal muscles using ultrasound and force sensor

低肌肉质量是一种普遍的身体组成异常在老年人和人与急性和慢性疾病。低肌肉质量是不良健康结果的一个强有力的预测指标，包括死亡率、术后并发症、住院或重症监护病房住院时间、代谢紊乱、身体和认知功能障碍以及生活质量差。除了对个人健康风险进行分层外，身体成分评估还可以为个性化的营养和运动干预提供信息，并评估治疗效果。超声是一种新兴的低成本、无创、便携的成像方式。超声可以准确测量人体不同部位骨骼肌的横截面积、厚度和回波强度。这些特点使超声成为临床实践中使用的最佳选择，而其他技术将受试者暴露在辐射下，需要分配空间，并且价格昂贵（例如，计算机断层扫描，双能x射线吸收仪）。然而，操作员对组织压缩不一致和手动图像分割等局限性仍然阻碍了其广泛应用。针对直接的临床需求，我们建议开发和测试(1)一种力传感器，用于标准化超声传感器的力，从而标准化组织压缩；(2)利用机器学习识别组织特征并自动评估肌肉参数的深度学习自动化算法。为了实现这一目标，我们与阿尔伯塔大学人类营养研究部门的研究团队建立了战略合作伙伴关系，该部门是加拿大最先进的身体成分评估研究机构，也是医学成像和生物医学工程的最佳研究机构。我们希望开发一种技术解决方案，弥合研究和技术验证之间的差距，以便医疗保健提供者可以广泛使用超声波进行肌肉质量评估，以改善可预见的未来的健康结果。

Low muscle mass is a prevalent body composition abnormality among older adults and people with acute and chronic conditions. Low muscle mass is a strong predictor of adverse health outcomes, including mortality, postoperative complications, length of hospital or intensive care unit stay, metabolic disorders, physical and cognitive dysfunction, and poor quality of life. In addition to stratifying individual health risks, body composition assessment also informs personalized nutrition and exercise interventions and evaluates treatment efficacy. Ultrasound is an emerging low-cost, non-invasive, portable, imaging modality. Ultrasound can accurately measure the cross-sectional area, thickness, and echo intensity of skeletal muscles at diverse body sites. These features make ultrasound an optimal choice for use in clinical practice compared to other techniques that expose subjects to radiation, need an allocated space, and are expensive (e.g., computed tomography, dual-energy x-ray absorptiometry). However, limitations, such as inconsistent tissue compression across operators and manual image segmentation, still preclude its widespread use. Responding to a direct clinical need, we propose to develop and test (1) a force sensor to standardize ultrasound transducer force and, consequently, tissue compression; and (2) a deep learning automated algorithm using machine learning to identify tissue features and automatically assess muscle parameters. To achieve this, we strategically established a partnership between research teams from the University of Alberta Human Nutrition Research Unit, a state-of-art research facility for body composition assessment and the best in Canada, medical imaging, and biomedical engineering. We hope to develop a technical solution for bridging the gap between research and technology validation so that healthcare providers can widely use ultrasound for muscle mass assessment to improve health outcomes in the foreseeable future.