Exploring advanced FDG PET-CT imaging feature analysis for more accurate diagnosis and outcome prediction in suspected large vessel vasculitis

这个项目的目的是发现和验证从有和没有左心室患者的FDG PET-CT图像中提取的预测参数，这些参数将允许开发能够有效地确定左心室患者预后的放射组学模型。目标：-利用基线FDG PET-CT数据整理大型单中心队列患者的结果数据-使用先进的特征分析和机器学习在测试和验证队列中开发假定的诊断和预测成像生物标记物，从来自单个中心的数据集的具有和不具有GCA的患者的FDG PET-CT图像-将结果与提取的参数相关联，以确定哪些参数具有最佳性能特征-验证多中心FDG PET-CT成像队列中的测试性能，允许协调不同扫描仪的过程。作为英国已建立的GCA多中心试验(TARGET Consortium，MRC)的一部分促进的数据池--探索这一过程在(A)其他形式的LVV中的潜力，和(B)MRA和CTA成像用于该项目的设施将包括：-用于构建放射学模型的深度学习软件-图像分析软件(Lifex)-2011-2017利兹LVV队列的基线FDG PET-CT数据-FDG PET-CT，来自接受测试但未被诊断为使用其他中心的LVV-FDG PET-CT数据的患者，以验证2011-2017利兹LVV队列治疗期间获取的模型-MRA和CTA数据医学图像分析需要对医疗设备/医疗设备与疾病之间的相互作用有所了解。虽然对左心室的PET-CT、MRA和CTA成像已经很好地建立了，但这些扫描提供的数据并不都会被分析；与LVV预后相关的放射学特征(提取的参数)目前尚不清楚，但可能包括：-血管的异质性(纹理分析)-炎症的大小、形状和位置-与邻近组织的关系(需要考虑诸如能量和熵等参数的统计分析)-信号密度-组织的信号密度-可能是疾病特有的、先前未确定的参数。一旦确定了参数，就可以建立它们与疾病及其表型的关系。这些特征将容易受到数据的图像质量和图像处理方法的影响，例如在纹理分析的情况下过度平滑。一旦建立了参数，我们将创建一种新的放射组学模型，它可以：(A)确定给定图像中的预测性生物标记物，(B)做出患者结果的预测。为此，我们将在卷积人工神经网络的帮助下利用深度学习。这个网络的多个层将服务于一个目的，从数据提取到对给定特征的评估。图像将层叠通过标准加权的层，模型将对预测的准确性和精确度做出判断。该模型将需要经过训练，以识别具有大量数据集的预测生物标记物，以最大限度地减少错误预测。由于该项目将由处于LVV研究前沿的临床医生共同监督，因此它将被设计为具有临床相关性和实用性。

The aim of this project is to discover and validate predictive parameters extracted from FDG PET-CT images of patients with and without LVV that will allow the development of a radiomic model that can determine the prognosis of LVV patients effectively.Objectives:- Collate outcome data for a large single centre cohort of patients with baseline FDG PET-CT data- Use advanced feature analysis and machine learning to develop putative diagnostic and predictive imaging biomarkers in test and validation cohorts derived from FDG PET-CT images of patients with and without GCA from dataset from a single centre - Correlate outcomes with extracted parameters to determine which parameters have the best performance characteristics- Validate test performance within a multi-centre FDG PET-CT imaging cohort allowing for harmonisation of the process across different scanners. Data pooling facilitated as part of an established GCA multi-centre trial in the UK (TARGET Consortium, MRC)- Explore the potential of this process in (a) other forms of LVV, and (b) MRA and CTA imagingThe facilities used for this project will include:- Deep learning software to build a radiomic model- Image analysis software (LifeX)- Baseline FDG PET-CT data from the 2011-2017 Leeds LVV cohort - FDG PET-CT from patients who were tested but not diagnosed with LVV- FDG PET-CT data from other centres to validate the model - MRA & CTA data taken during the treatment of the 2011-2017 Leeds LVV cohortThe analysis of medical images requires developing an understanding of interactions between modality / modalities and disease. While PET-CT, MRA and CTA imaging of LVV is well established, not all the data provided by these scans will have been analysed; the relevant extracted features could provide valuable insights into the characteristics of the disease that have not been discovered through routine / traditional clinical approaches.The radiomic features (the parameters extracted) relevant to LVV prognosis are currently unknown but could include:- the heterogeneity of the vessel (textural analysis)- size, shape and location of inflammation- relationship with neighbouring tissues (requires statistical analysis considering parameters such as energy and entropy)- intensity of signal- density of tissue- parameters that may be unique to the illness and not established previouslyOnce the parameters are determined, their relationship to the disease and its phenotype may be established. The features will be vulnerable to image quality of the data and image processing methods such as over smoothing in the case of textural analysis. Therefore the methods will require optimisation and will not necessarily be directly transferable from previous work.Once the parameters have been established, we will create a novel radiomic model that can: (a) determine predictive biomarkers in a given image and (b) make patient outcome predictions. For these purposes, we will utilise deep learning with the help of a convolutional artificial neural network. The multiple layers of this network will serve a purpose, ranging from data extraction, through to the evaluation of a given feature. The image will cascade through the layers being weighted by the criteria and the model will make a judgement about the accuracy and precision of the prognosis. The model will need to be trained to recognise the predictive biomarkers with a large data set to minimise incorrect predictions.As the project will be co-supervised by clinicians at the forefront of LVV research, it will be designed to be clinically relevant and useful.