Artificial Intelligence for Infectious Disease Imaging

来自综合研究机构（IRF）SARS-CoV-2动物成像研究的非人类（NHP）CT数据和来自COVID-19大流行的人类CT数据的可用性为探索NHP和人类CT肺部病变数据的相互作用提供了机会，以增强具有相同感染的受试者的深度学习（DL）图像分割能力。为了提高使用深度学习医学成像分割（DMIST）管道开发的DL模型的准确性，我们一直致力于利用可用的NHP数据来改进肺部病变分割。我们正在探索不同的方法，例如开发NHP预训练模型和深度学习跨物种迁移学习。 对此的实验已经显示出有希望的结果，其中预训练的模型和在NHP和人类CT数据的组合上训练的模型优于仅在人类数据上训练的模型。 这是一个激励，因为IRF-Frederick拥有大量的动物成像数据集，可以利用这些数据集来提高分割模型的性能。 目前，有多个卷积神经网络（CNN）框架已在DMIST管道中成功实现，并导致利用DMIST管道工作流程来调整和探索模型参数，以改善NHP和人类CT数据中的肺部病变分割。在对DL分割模型进行改进的同时，向DMIST管道添加模型比较模块的努力有助于扩展DMIST功能和工作流程。该模块的添加允许几乎没有编程知识的用户能够利用简单的配置文件来比较不同的模型参数。这有助于IRF-Frederick的研究人员不仅快速完成模型开发过程，还可以通过快速比较单个基准数据集上的任何数量的模型并汇总结果来快速完成评估过程。 监测用于医学成像应用的新的最先进的（SOTA）开发的CNN模型可以显着减少内部开发这些模型的工作。目前，U-NET、U-NET Transformer（UNET-R）和移位窗口U-NET Transformer（Swin-UNETR）已在DMIST管道中实现，并已成功用于SARS-CoV-2受试者的肺部病变分割。然而，由于与收集注释相关的挑战，肺部病变模型的预测准确性仍然缺乏。我们探讨了nnUNET模型架构，该架构已被证明是成功的各种各样的分割任务的医学图像上的NHP肺部病变CT数据的模型进行训练。我们看到，这种SOTA模型架构优于Swin-UNETR。这表明SOTA DL分割模型可以成功地用于各种分割任务，但应进行研究以找到解决IRF-Frederick相关图像分割需求的最佳解决方案。 人工智能团队已经建立了实验室的软件基础设施，使未来的分析能够更快，更一致地完成，而无需编辑基本代码。特别是，对深度学习医学图像分割工具包模型训练管道（DMIST-Train）进行了重大改进。添加了新功能，包括交叉验证的模型训练，其他模型和可编辑参数。添加的一个主要新功能是从配置文件配置整个培训过程的能力。这允许不了解管道内部工作原理的人运行深度学习模型训练并使用超参数进行实验。另一个主要的新功能是训练自我监督模型的能力，这是一种处于深度学习研究前沿的新兴技术。这些新功能中的每一个都需要对代码进行重大修改，但都是以模块化的方式完成的，这样未来的新功能就更容易了。AI团队还构建了一个全新的工具包来运行经典机器学习。这些是基于之前编写的脚本来运行SARS-CoV-2分类研究的实验，但经过重建以适用于任何新结构数据集。通过将脚本重建为模块，它使我们能够扩展我们可以运行的实验范围，以包括交叉验证循环中的回归模型训练和特征选择。重建还意味着添加到工具包中的新图和功能可以轻松应用于增强多个数据集的所有研究。 深度学习医学图像分割工具包模型部署管道（DMIST-Deploy）也做出了重大贡献。具体贡献包括通过创建conda和pip需求/规范文件来标准化Python环境设置过程，从NIFTI文件类型实现放射组学特征的批量计算，以及向放射组学特征计算的输出添加额外的元数据细节。优化参数或执行参数网格搜索时生成多个训练配置文件的过程是自动化的。为了提高代码质量，通过将标准Python格式化模块“black”应用于代码库，改进存储库README文件，并在整个源代码中添加有用的注释，对代码可读性和文档进行了改进。此外，AI团队还帮助开发和实施了GitHub上协作软件开发的最佳实践。

The availability of non-human (NHP) CT data from SARS-CoV-2 animal imaging studies at the Integrated Research Facility (IRF) and human CT data from the COVID-19 pandemic has presented an opportunity to explore the interaction of combining NHP and human CT lung lesion data to enhance deep learning (DL) image segmentation capabilities for subjects with the same infection. To enhance DL model accuracies developed with the Deep learning Medical Imaging SegmenTation (DMIST) pipeline, we have been working on leveraging available NHP data to improve lung lesion segmentation. We are exploring different methodologies such as developing NHP pretrained models and deep learning inter-species transfer learning. Experimentation of this has shown promising results where pretrained models and models trained on a combination of NHP and human CT data outperform models which are only trained on human data. This is motivating as the IRF-Frederick has a large collection of animal imaging datasets that could be leveraged to improve segmentation model performances. Currently, there are multiple convolutional neural network (CNN) frameworks have been successful implemented in the DMIST pipeline and has led to utilizing the DMIST pipeline workflow to tune and explore model parameters for improving lung lesion segmentation in both NHP and human CT data. In-line with making improvements to DL segmentation models, efforts to add a model comparison module to the DMIST pipeline has helped expand DMIST capabilities and workflows. The addition of the module allows users with little to no programming knowledge to be able to compare different model parameters to one another utilizing a simple configuration file. This helps enable IRF-Frederick researchers to quickly iterate through not only the model development process but also the evaluation process by quickly comparing any number of models all on a single benchmark dataset and aggregate results. Monitoring the new state-of-the-art (SOTA) developed CNN models for medical imaging applications could significantly reduce efforts to develop these models inhouse. Currently, the U-NET, U-NET Transformer (UNET-R), and Shifted Window U-NET Transformer (Swin-UNETR) are implemented in the DMIST pipeline and have been successful for lung lesion segmentation in SARS-CoV-2 subjects. However, the predictive accuracies of the lung lesion models are still lacking due to challenges related to collecting annotations. We explored the nnUNET model architecture which has been shown to be successful for a wide variety of segmentation tasks on medical images by training the model on NHP lung lesion CT data. We see that this SOTA model architecture is outperformed by the Swin-UNETR. This demonstrations SOTA DL segmentation models can be successful for various segmentation tasks but it should be investigated to find the optimal solution to solve IRF-Frederick related image segmentation needs. The AI team has built up the software infrastructure of the lab, enabling future analyses to be completed faster, more consistently, and without the need to edit the base code. Particularly, major improvements were made to the Deep-learning Medical Image Segmentation Toolkit model Training pipeline (DMIST-Train). New capabilities were added including cross-validated model training, additional models, and editable parameters. One major new capability that was added was the ability to configure the entire training process from a configuration file. This allows people without knowledge of the inner workings of the pipeline to run deep learning model training and experiment with hyperparameters. Another major new capability is the ability to train a self-supervised model, a burgeoning technique that is at the forefront of deep learning research. Each of these new capabilities required a significant overhaul of the code but was done in a modular way such that future new additions will be easier. The AI team also built an entirely new toolkit for running classical machine learning. These were based on previously written scripts to run the experiments for the SARS-CoV-2 Classification study but were rebuilt to be applicable to any new structure dataset. By rebuilding the scripts into modules, it enables us to expand the scope of experiments we could run to include regression model training and feature selection within the cross-validation loop. The rebuild also means that new plots and capabilities added to the toolkit can be easily applied to enhance all studies across multiple datasets. Significant contributions were also made to the groups Deep-learning Medical Image Segmentation Toolkit model deploy pipeline (DMIST-Deploy). Specific contributions include standardizing the Python environment setup process by creating conda and pip requirements/specification files, implementing bulk calculation of radiomic features from NIFTI filetypes, and adding additional metadata details to the output of radiomic feature calculation. The process for generating multiple training configuration files when optimizing parameters or performing a parameter grid search was automated. To enhance code quality, improvements were made to code readability and documentation by applying the standard Python formatting module "black" to the codebase, improving repository README files, and adding useful comments throughout the source code. Additionally, the AI team helped develop and implement best practices in collaborative software development on GitHub.