Synergistic Image Analysis of longitudinal cardiac MRI

博士项目的目标开发和评估框架，以改善心脏生物标志物的估计及其从纵向磁共振成像的变化研究技术范围从单独的独立分析到“协同”和同时分析多个时间点的扫描建立心脏生物标志物估计的可重复性使用重复扫描数据将纵向分析结果与基线可重复性性能进行比较项目描述：近年来已经看到最先进的深度学习分割工具的出现，在一系列任务上达到人类水平的表现。在心脏MR中，这样的技术已经发展到获得成像生物标志物，用于全自动质量控制的功能量化[2]。在临床实践和研究中，对受试者进行多次扫描是很常见的。由此产生的纵向数据将包含显著的相似性，以及由于解剖和病理随时间变化而产生的差异，以及由于获取差异而产生的差异。然而，用于分割结构和从成像数据中提取生物标志物的图像分析工具通常不会利用这些相似性，而是将数据视为独立的。我们需要能够灵敏地识别生物标志物变化的方法，特别是在试验一种新药或治疗方法时，因为这将使试验规模/持续时间减少，这可能对成本和确定技术有效性的能力产生重大影响。其他一些研究领域，如图像重建，已经开发了“协同”方法来处理纵向数据[3]。在图像分析方面也进行了一些相关工作，但迄今为止，这主要集中在利用纵向数据进行分类问题，如预测治疗反应[4,5]、肿瘤检测[6,7]或评估疾病严重程度[bb0]。该项目将寻求开发纵向成像数据的“协同”分析框架，用于生物标志物估计的新应用。我们的重点将放在心脏磁共振成像上，但我们相信所开发的方法将适用于更广泛的模式和问题，例如，所开发的方法可以扩展到更多的器官，以评估多器官疾病（如Covid-19）的健康状况。我们的目标是以协同的方式利用扫描之间的相似性，以便对所有扫描的形态和功能生物标志物进行更准确和稳健的估计。我们还将使用来自同一受试者的重复扫描数据来建立可重复性/再现性水平，并评估我们针对这一点的协同方法。首先，我们将开发用于协同分析同一扫描仪在不同时间点的两次扫描的方法，但随后寻求将框架推广到多个时间点的多个扫描仪的多次扫描。我们将主要关注通过分割提取生物标志物，但也将研究从成像数据中直接协同估计生物标志物的可能性。

Aim of the PhD Project Develop and evaluate frameworks for improving the estimation of cardiac biomarkers and their changes from longitudinal MR imaging Investigate techniques ranging from separate independent analysis to 'synergistic' and simultaneous analysis of scans from multiple time points Establish the repeatability of cardiac biomarker estimation using repeat scan data Compare longitudinal analysis results to baseline repeatability performance Project description:Recent years have seen the emergence of state-of-the-art deep learning segmentations tools, reaching human-level performance on a range of tasks [1]. In cardiac MR, such techniques have been developed to derive imaging biomarkers for fully automated quality-controlled functional quantification [2]. In both clinical practice and research studies, it is common for subjects to be scanned multiple times. The resulting longitudinal data will contain significant similarities, as well as differences due to changes in anatomy and pathology over time as well as variations due to differences in acquisition. However, the image analysis tools that are employed to segment structures and derive biomarkers from imaging data do not typically exploit these similarities and treat the data as if they were independent. There is a need for methods that can sensitively identify biomarker changes, especially when trialling a new drug or treatment, as this would enable the trial size/duration to be reduced, which can have a significant impact on cost and the ability to determine the efficacy of the technology. Some other strands of research, such as image reconstruction, have developed 'synergistic' approaches for processing longitudinal data [3]. Some related work has also been performed in image analysis, but to date this has mostly focused on exploiting longitudinal data for classification problems such as predicting treatment response [4,5], tumour detection [6,7] or assessing disease severity [8]. This project will seek to develop a 'synergistic' analysis framework for longitudinal imaging data in a novel application to biomarker estimation. Our focus will be on cardiac MR imaging, but we believe that the methods developed will be applicable to a wider range of modalities and problems, for example the methods developed may be extended to further organs to assess health in multi-organ conditions such as Covid-19. We will aim to exploit the similarities between scans in a synergistic way to enable more accurate and robust estimates to be made of morphological and functional biomarkers from all scans. We will also use repeat-scan data from the same subjects to establish a level of repeatability/reproducibility and evaluate our synergistic approaches against this. To begin with, we will develop methods for synergistically analysing two scans from the same scanner at different time points, but then seek to generalise the framework to multiple scans from multiple scanners at multiple time points. We will predominantly focus on deriving biomarkers via segmentations but will also investigate the possibility of synergistic direct estimation of biomarkers from the imaging data.