General purpose abnormality detection in whole body PET

博士项目目标：开发人工智能方法，可以从解剖数据中预测健康的PET放射性示踪剂摄取。使用贝叶斯深度学习以异方差（空间变化）的方式对这些预测的不确定性进行建模。将这些算法转化为临床实践，并将其整合到临床工作流程中。项目描述/背景：癌症在正电子发射断层扫描（PET）中可能具有高度异质性的异常模式。这些特殊的异常模式对于检测、分期和预测疾病的发展非常重要。在研究背景下，图像通常通过执行组比较来分析。这种方法不符合临床情况，在临床情况下，分析必须在个人水平上进行，以检测特定于受试者的模式。在临床实践中，PET图像大多是视觉分析。这种方法的敏感性和特异性在很大程度上取决于观测者的经验，不适合无法获得高级图像读取专业知识的中心[Perani等人，2014]。PET图像的定量分析将通过帮助确定正常和病理结果之间的客观界限来缓解这一问题。PET摄取可以在区域或逐体素的基础上进行定量评估。在区域分析中，将区域摄取与正常对照人群的预期区域摄取进行比较。这种分析通常需要先验知识来选择合适的图谱和相关的判别区域，这些区域应该适应特定的病理，限制其使用（Signorini等人，2019）。在体素分析中，受试者的PET图像通常与标准化群体空间对齐，以逐体素为基础，将空间规范化扫描的代谢活动与正常对照扫描获得的分布进行比较。在Neurostat等软件中实现的方法包括将被调查对象的PET图像注册到标准空间，并通过z分数将其与对照人群进行比较。然后将z分数地图投影到不同的表面上，从而产生用于图像解释的三维立体定向表面投影。其他实现类似技术的软件工具已被用于PET数据的分析，例如GE Healthcare的NeuroGam。这些软件包主要是为脑成像开发的，限制了它们对其他身体疾病的适用性，但也有一些关于异常统计模型的非最佳假设（噪声模型，参数分布等）。博士候选人将致力于使用人工智能语义回归模型，以最佳方式为全身数据创建健康PET示踪剂分布的患者和示踪剂特定模型（Klaser et al. 2019）。由于预期示踪剂分布的非高斯性质，这些模型将使用基于深度学习的不确定性估计，使用Dropout (Gal et al. 2015)，结合多个假设输出预测（M-heads）来创建示踪剂分布的鲁棒统计模型。由于这种模型的抽样性质，然后可以使用非参数统计检验来估计PET信号的每像素异常程度（Burgos等人，2017），使得该技术在临床环境中的应用是安全的。此外，由于PET成像具有局部和全局摄取模式，因此必须对架构进行优化，以考虑到全身语义，以便在多个空间尺度上适当地模拟PET。这些模型将应用于一大批进行全身PET成像的癌症患者，无论是PET- ct还是PET- mri。

Aim of the PhD Project:Develop AI methods that can predict healthy PET radiotracer uptake from anatomical data.Model the uncertainty of such predictions in a heteroscedastic (spatially varying) manner using Bayesian deep learning.Translate these algorithms to clinical practice and integrate them within the clinical workflow.Project Description / Background:Cancers can have a highly heterogeneous pattern of anomalies in positron emission tomography (PET). These specific patterns of anomaly are important to detect, stage and predict the evolution of disease. In a research context, images are often analysed by performing group comparisons. This approach does not correspond to the clinic scenario, where the analysis has to be performed at the individual level to detect subject-specific patterns. In clinical practice, PET images are mostly analysed visually. The sensitivity and specificity of this approach greatly depends on the observer's experience and is not in favour of centres where advanced expertise in image reading is unavailable [Perani et al., 2014]. Quantitative analysis of PET images would alleviate this problem by helping define an objective limit between normal and pathological findings.PET uptake can be quantitatively evaluated either regionally or on a voxel-by-voxel basis. In regional analysis, the regional uptake is compared with the regional uptake expected in a normal control population. This analysis usually requires prior knowledge to select the appropriate atlas and relevant discriminant regions, which should be adapted to a specific pathology, limiting its use (Signorini et al., 2019).In voxel-wise analysis, a subject's PET image is usually aligned to a standardised group space to compare the metabolic activity of the spatially normalised scan to a distribution obtained from normal control scans, on a voxel-by-voxel basis. The approach implemented in software such as Neurostat consists of registering the PET image of the subject under investigation to a standard space and comparing it to a population of controls by means of a Z-score. The Z-score map is then projected onto different surfaces resulting in three-dimensional stereotactic surface projections that are used for image interpretation. Other software tools implementing a similar technique have been used for the analysis of PET data, such as NeuroGam by GE Healthcare. These packages have been mostly developed for brain imaging, limiting their applicability for other bodily diseases, but also have several non-optimal assumptions (noise model, parametric distribution, etc) about the statistical models of abnormality.The PhD candidate will work towards creating patient- and tracer-specific models of healthy PET tracer distribution in an optimal way for full body data using artificial intelligence semantic regression models (Klaser et al. 2019). Due to the non-Gaussian nature of the expected tracer distribution, these models will use deep learning based uncertainty estimation using Dropout (Gal et al. 2015), jointly with multiple hypothesis output predictions (M-heads) to create a robust statistical model of tracer distribution. Due to the sampling nature of such model, a non-parametric statistical test can then be used to estimate a per-pixel degree of abnormality of the PET signal (Burgos et al. 2017), making the application of this technique safe in a clinical setting. Further to this, as PET imaging has both local and global uptake patterns, architectures will have to be optimised to take the full body semantics into account as to appropriately model PET at multiple spatial scales. These models will be applied to a large cohort of cancer patients with full body PET imaging, either in a PET-CT or PET-MRI setting.