Towards building advanced machine learning image translation models to estimate Amyloid-Beta and Tau PET images from structural MRI

图像翻译是机器学习的一种高级形式。它可以用来从低成本的输入中生成高成本的图像。阿尔茨海默氏症是最常见的痴呆症类型，影响着全球数千万人。一种流行的假说认为，淀粉样β蛋白和牛磺酸是阿尔茨海默病的主要原因，它们聚集在大脑中，形成众所周知的“斑块”。这些斑块在死后痴呆患者的大脑显微镜上已经有一百多年的历史了，最近可以在活体内用PET成像进行可视化。然而，与传统的结构磁共振成像(500美元)相比，PET成像是一项非常昂贵的测试(3000至8000美元)。它还具有更大的侵入性，需要在手术过程中注射辐射诱导示踪剂和采血。此外，并非所有司法管辖区都有示踪剂和技术(在艾伯塔省的使用非常有限！)。MRI的优点是更实惠和更容易获得，它的侵入性更小，但它不能提供与PET数据相同的详细分子信息。先前的研究表明，淀粉样β蛋白和Tau的PET和结构MRI之间存在相互的信息。该项目将利用机器学习将结构MRI图像转换为淀粉样β蛋白和Tau PET图像。关键的前期工作已经完成，包括收集所需的数据、获得机构伦理批准和建立2D概念验证模型。目标1：开发一个简单的可用于PET估计的3D U-Net模型；目标2：利用代价函数和3D GaN改进模型性能；目标3：实现Vision Transformer GaN并与CNN-GaN进行比较。从基础核磁共振获得淀粉样β蛋白和Tau PET图像的影响对于阿尔茨海默病的早期检测非常重要。这将促进一种筛查方法，因为重复MRI是非常实用和安全的，而重复PET与辐射暴露和注射辐射诱导示踪剂有关。这是一个值得注意的有价值的命题，曾经经过提炼和验证。定量图像翻译的进展可以应用于许多其他高价值的应用。

Image-translation is an advanced form of machine learning. It can be used to generate high-cost images from lower-cost inputs. Alzheimer's Disease is the most common type of dementia, impacting tens of millions of people worldwide. A popular hypothesis proposes Amyloid-Beta and Tau are the main causes of Alzheimer's Disease, aggregating in the brain forming the well-known 'plaques.' These plaques have been visible on microscopes from the brains of post-mortem demented patients for over a hundred years, and more recently can be visualized with PET imaging in-vivo. PET imaging is, however, a very expensive test to run ($3000-to-$8000), relative to a conventional structural MRI ($500). It is also more invasive requiring injection of a radiation inducing tracer and blood sampling during the procedure. Furthermore, the tracers and technologies are not available in all jurisdictions (Very limited use in Alberta!). MRI has the advantage of being more affordable and accessible, it is less invasive, however it does not provide the same detailed molecular information as PET data. Previous studies indicate there is mutual information relating Amyloid-Beta and Tau PET and structural MRI. This project will utilize machine learning to translate structural MRI images to Amyloid-Beta and Tau PET images. Key preliminary work has already been completed, including gathering of the required data, obtaining institutional ethics approval, and building a 2D proof-of-concept model. Aim 1: Develop a Simple 3D U-Net Model that is Capable of PET Estimation; Aim 2: Improve Model Performance with Cost Functions and 3D GAN; Aim 3: Implement Vision Transformer GAN and compare with CNN-GAN. The impact of obtaining Amyloid-Beta and Tau PET images from basic MRI is important for the early detection of Alzheimer's disease. This would facilitate a method for screening, as repeat MRI is very practical and safe, while repeat PET is associated with radiation exposure and the injection of radiation inducing tracer. A notable valuable proposition once refined and validated. Advances in quantitative image translation could be applied to many other high value applications.