Artificial Intelligence for Infectious Disease Imaging

CNN架构用于创建基于深度学习的肝脏分割模型，该模型在非人类灵长类动物（NHP）的CT扫描上看到。CNN架构通过实现特征金字塔网络（FPN）进行了优化。然后利用该FPN模型将肝脏分割推广到其他器官，如脾脏，然后在SARS-CoV-2感染的NHP模型中推广到整个肺和肺部病变，如在IRF进行的那样。为了更好地与COVID-19的组织病理学发现相关联，开发了肺叶分割。这是一项正在进行的工作，因为CT扫描上的肺叶注释数量最少，因为难以创建此地面实况数据。目前，正在探索使用少量训练数据的方法。 为了使这些基于深度学习的图像分割方法可供IRF的成像社区使用，开发了一个自动化的流水线过程。该流水线利用NIAID高性能计算环境（Locus）来允许使用图形处理单元（GPU）进行增强的并行处理。 IRF的其他成像研究人员正在使用这些方法对SARS-CoV-2、马尔堡、埃博拉、尼帕和拉沙病毒的传染病成像研究期间获得的图像进行后处理。此外，这些分割提供了用于放射组学特征提取的感兴趣区域。这数百个放射组学特征已被输入到特征选择方法中，例如最大相关最小冗余（mRMR），以减少和优化用于预测分析的特征。此外，开发了一种新的特征选择方法（mRMR-permute），该方法使用排列测试来自动限制所选特征的数量。预测分析是用传统的机器学习方法实现的，这些方法被探索以确定SARS-CoV-2成像实验的最佳用途。 为了进一步提高图像分割的特异性，正在开发病变表型分类算法。作为初步实施，CT扫描中看到的肺部病变将被分类为磨玻璃样阴影或实变。这种分类将有助于纵向量化期间的严重程度评估。

A CNN architecture was used to create a deep learning-based segmentation model of the liver seen on CT scans of a nonhuman primate (NHP). The CNN architecture was optimized by implementing a feature pyramid network (FPN). This FPN model was then utilized to generalize liver segmentation to other organs such as the spleen and then the whole lung and lung lesions in NHP models of SARS-CoV-2 infection, as conducted at the IRF. Lung lobe segmentation was developed in order to better correlate with histopathologic findings of COVID-19. This is a work-in-progress as the number of lung lobe annotations on CT scans is at a minimum due to the difficulty in creating this ground truth data. Currently, methods that function with low numbers of training data are being explored. To make these deep learning-based image segmentation methods available for use by the imaging community at the IRF, an automated pipeline process was developed. This pipeline utilizes the NIAID high performance computing environment (Locus) to allow for enhanced parallel processing with graphical processing units (GPUs). These methods are being used by other imaging researchers at IRF to post-process images acquired during infectious disease imaging studies of SARS-CoV-2, Marburg, Ebola, Nipah and Lassa viruses. In addition, these segmentations provide regions of interest for radiomic feature extraction. These hundreds of radiomic features have been input into feature selections methods, such as maximum relevance minimum redundancy (mRMR), to reduce and optimize features for predictive analyses. In addition, a novel feature selection method was developed (mRMR-permute), which uses permutation testing to automatically limit the number of features chosen. The predictive analyses are implemented with conventional machine learning methods that were explored to determine optimal use for SARS-CoV-2 imaging experiments. To further the specificity of image segmentations, a lesion phenotype classification algorithm is being developed. As an initial implementation, lung lesions seen in CT scans are to be classified as ground glass opacities or consolidations. This classification will help with severity assessment during longitudinal quantification.