Machine Learning on 3D Ultrasound Images and Wearable IoT Data for Brace Treatment of Spinal Deformities

支持支具控制脊柱畸形有效性的科学尚不清楚。脊柱对物理载荷的机械响应无法实时检查。本研究旨在开发、推进和整合监测技术，以调查和了解模拟支具施加的生物力学载荷如何影响脊柱内部对线。将研究四个工程挑战：1）创建低成本3D便携式无线超声设备，2）开发机器学习（ML）算法以从超声图像中自动提取3D参数，3）开发软件平台以记录施加到身体的压力的负载，位置和方向，并通过实时3D超声与内部对齐变化相关联，以及4）设计和开发可穿戴物联网（IoT）设备以监测支具使用并控制支具和患者身体之间的界面压力，以了解如何在个体基础上优化脊柱对齐。 一个低成本的2D便携式无线超声（US）设备集成了电磁位置和方向跟踪系统将被增强，以创建一个3D US扫描仪。重建过程的精度和速度对于实时应用是重要的。 将开发基于卷积神经网络的ML算法，以自动从超声图像中提取参数。 将使用300张US图像进行培训，100例用于测试，100例用于确认。 这种方法可以减少人为测量误差，并允许非超声专家进行培训以提取信息。 将开发一个与无线传感网络集成的软件平台，以实时调查不同载荷下的脊柱内部对准变化。内置于定制站立框架内的无线压力控制传感器网络已嵌入压力垫中，以跟踪空间和压力信息。所开发的便携式3D US将用于捕获内部对准变化，同时操纵方向、位置和来自压力垫值的测量压力。这个创新的工具帮助矫形师在支具设计过程中获得实时反馈。 物联网设备由微计算机系统、压力传感器、嵌入压力区域下的气囊、泵和阀反馈系统、温度传感器和方位传感器组成，将设计和建造用于在日常生活中监测支具佩戴时间和控制佩戴支具的松紧度。 该设备可以根据个人需求进行编程，以优化支具的有效性。 这项研究的结果将是一套工具，可以在没有电离辐射的情况下实时成像脊柱内部对准。ML算法可以应用于具有不同训练集的其他医学成像应用。可穿戴物联网设备可以用于许多其他健康监测系统。

The science supporting the effectiveness of bracing to control spinal deformity is not clear. The mechanical response of the spine to physical loadings cannot be examined in real-time. This research aims to develop, advance and integrate monitoring technologies to investigate and understand how the biomechanical loadings applied from a simulated brace affects the internal spinal alignment. Four engineering challenges will be investigated:1) creation of a low cost 3D portable wireless ultrasound device, 2) development of machine learning (ML) algorithms to automatically extract 3D parameters from ultrasound images, 3) development of a software platform to record loads, positions and directions of pressure applied to a body and correlate with the internal alignment changes via real-time 3D ultrasound, and 4) design and development of wearable Internet-of-Things (IoT) devices to monitor the brace usage and control the interface pressure between the brace and patient's body to understand how spinal alignment may be optimized on an individual basis. A low cost 2D portable wireless ultrasound (US) device integrated with an electromagnetic position and orientation tracking system will be enhanced to create a 3D US scanner. The accuracy and the speed of the reconstruction process are important for real-time applications. ML algorithms based on convolutional neural networks will be developed to automatically extract parameters from ultrasound images. 300 US images will be used for training, 100 cases for testing and 100 cases for validation. This approach can reduce human measurement errors and training to allow non-ultrasound experts to extract information. A software platform integrated with a wireless sensing network to investigate the internal spinal alignment changes with different loads in real-time will be developed. A wireless pressure control sensor network built inside a custom standing frame has been embedded into the pressure pads to track spatial and pressure information. The developed portable 3D US will be used to capture internal alignment changes while manipulating the orientation, location and measured pressure from value of pressure pads. This innovate tool helps orthotists to obtain real-time feedback during brace design. An IoT device consists of a microcomputer system, a pressure sensor, air bladders embedded under the pressure area, a pump and valves feedback system, a temperature sensor and orientation sensor will be designed and built to monitor brace wear time and control wear brace tightness during daily living. This device can be programmed for individual needs to optimize brace effectiveness. The results of this research will be a suite of tools which can image the internal spinal alignment without ionizing radiation in real-time. The ML algorithms can be applied to other medical imaging applications with different training sets. The wearable IoT devices can be utilized for many other health monitoring systems.