Machine Learning and Medical Image Analysis for Point-of-Care Ultrasound Systems

在低收入和中等收入国家(LMIC)，由于缺乏有经验的超声师和良好扫描所需的专业知识，获得诊断超声(US)的机会受到阻碍。此外，传统的美国系统背负着高昂的设备成本负担，降低了它们在此类环境中使用的可行性。统计数据显示，在世界范围内，99%的孕产妇死亡发生在LMICs中，这表明在美国提供诊断的重要临床需求尚未得到满足[1]。最近最先进的技术进步已经设计出低成本的美国设备，显示出在护理点(POC)场景中的高潜力。同样，机器学习体系结构的进步现在提供了优于谓词解决方案的性能，并在2D美国图像和视频的计算机模式识别中开辟了可能的领域。在印度和非洲等国的POC US系统中部署US的自动计算机分析具有真正的潜力，从而解决技能危机和诊断需求。拟议的博士研究旨在开发和评估一种新的自动图像分析框架，该框架利用简化的US多次扫描扫描协议，为不熟悉超声的医护人员提供临床决策支持工具。2D US图像和视频具有丰富的时空特征和声学模式，研究将考虑如何提取这些特征并在机器学习体系结构中结合起来，以揭示重要的临床信息。该项目的第一个关键挑战将是将从文献和临床合作者那里获得的临床标准转化为机器学习框架。第二个挑战将是如何为选定的临床任务设计和实现适当的机器学习体系结构。深度学习方法最有可能是一种可行的方法。最后，将与英国和印度的临床合作伙伴合作进行可行性研究，以评估开发的方法并考虑其在实践中的潜在可用性。该研究属于EPSRC的“医疗保健技术”和“工程”研究主题。特别是，这项研究发展了“图像和视觉计算”、“人机交互”和“医学成像”三个子主题的工作。这项研究也符合EPSRC的重大挑战，即“转变社区健康和护理”和“优化治疗”。博士研究将与GCRF资助的CALOPUS项目(计算机辅助低成本护理点超声：EP/R013853/1)相关，该项目是牛津大学生物医学工程研究所(IBME)和纳菲尔德妇女和生殖健康系以及印度法里达巴德的转换健康科学和技术研究所(THSTI)的联合合作项目。

Access to diagnostic ultrasound (US) in low and middle income countries (LMICs) is impeded because of a lack of experienced sonographers and the expertise required to scan well. Furthermore, conventional US systems carry the burden of high equipment costs reducing their feasibility for use in such environments. Statistics show that worldwide, 99% of maternal deaths occur in LMICs indicating an important unmet clinical need to provide US-based diagnosis [1]. Recent state-of-the-art advances have engineered low-cost US devices which show high potential for use in point-of-care (POC) scenarios. Similarly, advancements in machine learning architectures now offer superior performance over predicate solutions and open realms of possibility in computer pattern recognition of 2D US images and video. There is a real potential for automated computer analysis of US to be deployed within POC US systems in countries such as India as well as in Africa, thus addressing the skills crisis and diagnostic need.The proposed doctoral research aims to develop and evaluate a novel automated image analysis framework that utilises a simplified US scanning protocol of multiple scanning sweeps to provide a clinical decision support tool for healthcare workers unfamiliar with ultrasound. 2D US images and video are rich in spatial-temporal features and acoustic patterns and the research will consider how these can be extracted and combined to good effect within machine learning architectures to reveal important pieces of clinical information. A first key challenge of this project will be to translate the clinical criteria, obtained from the literature and clinical collaborators, into a machine learning framework. A second challenge will be how to design and implement appropriate machine learning architecture for the chosen clinical tasks. A deep learning approach is most likely to be a feasible method. Finally, feasibility studies will be performed in collaboration with clinical partners in the UK and India to evaluate the developed methods and consider the potential usability in practice.The research falls within the EPSRC's 'Healthcare Technologies' and 'Engineering' research themes. In particular, the research develops work within 'Image and Vision Computing', 'Human-Computer Interaction', and 'Medical Imaging' sub-themes. The research also fits into the EPSRC Grand Challenges through 'Transforming Community Health and Care' and 'Optimising Treatment'.The doctoral research will be conducted associated with the GCRF-funded CALOPUS Project (Computer-Assisted Low-cost Point-of-Care Ultrasound: EP/R013853/1) which is a joint collaboration between the Institute of Biomedical Engineering (IBME) and Nuffield Department of Women's and Reproductive Health, University of Oxford, and the Translational Health Science and Technology Institute (THSTI), Faridabad, India.