Toward Automated Video Quality Assessment of Ultrasound

传统上，医学图像分析文献考虑基于图像的诊断（解释捕获的图像）。超声图像分析的不同之处在于对图像的关注较少，更多的是对视频处理的关注。超声成像也是一种交互式实时成像技术，可以廉价地重新拍摄视频。人类专家很快就学会了如何在视频不够理想或“不符合目的”时重拍视频。计算机还无法模仿人类的这种能力。该项目旨在研究基于深度学习的视频分析算法，以推进自动化超声视频质量评估，朝着更类似于人类专家行为的通用解决方案发展。目的和目标我们有两个大的真实世界数据集可用于这项研究，这将使我们能够从不同的角度来看待超声视频质量评估。第一个是PULSE数据集，这是一个大规模的多模态徒手数据集，包括超声视频、凝视跟踪数据和超声医师在进行胎儿筛查扫描时获得的探头运动数据。第二个数据集来自一项基本的妊娠超声研究，超声医师在两个地点按照由预定义的线性扫描组成的简单扫描方案（称为CALOPUS超声方案（CUP））获得数据。我们将使用PULSE多模态数据集来研究人类专家在实践中用于确定视频质量的标准。这将提供对需要嵌入视频QA的一般深度学习模型的计算标准的见解。我们将使用CALOPUS数据集来定义开发的视频QA模型的训练和测试数据。我们将从简单的要求开始，如“如果胎儿头部存在，视频质量良好”。然后，我们将继续讨论更一般的要求，如“如果所有解剖结构都足够清晰，视频质量就会很好”。然而，我们不想为每个任务构建定制的解决方案，因为这是乏味的，所以我们的深度学习设计将需要越来越“智能”，并将致力于在设计中利用一些最新的计算机视觉思想，以便理想的人工注释不需要训练（自我监督），模型学习可以推广到新的未见过的任务（领域适应），并且它们是理想的可解释的。因此，博士研究将逐步探索开发和测试一系列更通用的基于深度学习的超声视频质量评估方法。新颖性和影响这项研究将促进对医疗保健成像领域的自动视频分析算法的理解。自动超声视频质量评估将是一项变革性技术，可以简化超声检查，并使其成为高收入国家和中低收入国家更广泛的临床专业人员可以使用的技术。与EPSRC主题保持一致该项目属于EPSRC医疗技术、信息通信技术、人工智能和机器人研究领域

Project ContextTraditionally the literature on medical image analysis considers image-based diagnostics (interpreting a captured image). Ultrasound image analysis is different as there is less focus on the image, and more on video processing. Ultrasound imaging is also an interactive real-time imaging technique where video can be cheaply re-taken. Human experts quickly learn how to retake a video if it is sub-optimal or not "fit-for-purpose". Computers are yet to be able to mimic this human capability. This project seeks to investigate deep learning-based video analysis algorithms to advance automated ultrasound video quality assessment towards a generalisable solution more akin to human-like expert behaviour.Aims and ObjectivesWe have two large real-world datasets available for this research which will allow us to look at ultrasound video quality assessment from different perspectives. The first is the PULSE dataset which is a large-scale multi-modal freehand datasets of ultrasound video, gaze tracking data, and probe motion data acquired while sonographers perform fetal screening scans. The second dataset is from a basic pregnancy ultrasound study where sonographers have acquired data at two sites following a simple scanning protocol consisting of pre-defined linear sweeps (called the CALOPUS ultrasound protocol (CUP)).We will use the PULSE multi-modal dataset to study the criteria that human experts use to determine video quality in practice. This will provide insight into the computational criteria that will need to be embedded in a general deep learning model of video QA. We will use the CALOPUS dataset to define the training and test data for the developed video QA models. We will start with simple requirements like "the video is good quality if the fetal head is present". We will then move on to more general requirements such as "the video is good quality if all anatomical structures are clear enough". However, we do not want to build a bespoke solution for every task as this is tedious to do, so our deep learning designs will need to be increasingly "intelligent" and will aim to utilise some of the latest computer vision ideas in their design so that ideally manual annotation is not needed for training (self-supervision), models learn to be generalisable to new unseen tasks (domain adaptation) and they are ideally explainable. The doctoral research will thus progressively explore developing and testing a family of deep learning-based ultrasound video quality assessment methods that are more general.Novelty and ImpactThis research will advance understanding of automated video analysis algorithms with a domain focus on healthcare imaging. Automated ultrasound video quality assessment would be a transformational technology to simplify ultrasound and make it an accessible technology for a wider range of clinical professionals in high-income and low-and-middle-income countries.Alignment with EPSRC ThemesThis project falls within the EPSRC healthcare technologies, ICT and artificial intelligence and robotics research areas