Synergistic integration of deep learning and regularized image reconstruction for positron emission tomography

项目摘要/摘要 正电子发射断层扫描(PET)是一种广泛应用于肿瘤学的高灵敏度分子成像手段， 神经学和心脏病学，有能力通过观察活体内分子水平的活动 注射特定的放射性示踪剂。除了常用的F-18-FDG外，正在开发新的示踪剂 不断发展和研究，以确定各种疾病的具体途径。新型PET扫描仪 还通过利用飞行时间(TOF)信息，实现交互能力的深度， 扩大了立体角度覆盖范围。为了充分发挥新的PET示踪剂和扫描仪的潜力， 对发展先进的图像重建方法的需求越来越大。这笔赠款 应用提出了一种协同集成深度的正则化图像重建新框架 学习和正则化图像重建。新的框架是由以下方面的最新进展促成的 机器学习，它提供了一种工具来消化嵌入在现有医学中的大量信息 图像。该方法将预先训练好的深度神经网络嵌入到迭代图像重建中 框架，并直接使用深度神经网络对PET图像进行正则化。通过训练深层神经 网络中含有大量高质量的低噪声PET图像，该方法可以捕捉到复杂的图像 来自现有的受试者间和受试者内数据的先验信息，因此预计基本上 优于目前最先进的正则化图像重建方法。这样做的两个具体目标 探索性建议是(1)开发理论框架，以协同方式整合深度学习在 正则化正则化PET图像重建算法的实现与验证 使用现有动物数据的有效性。一旦使用现有的动物数据验证了所提出的方法， 我们会寻求拨款，以取得所需的人体数据，以便在 临床PET扫描仪。

Project Summary/Abstract Positron emission tomography (PET) is a high-sensitivity molecular imaging modality widely used in oncology, neurology, and cardiology, with the ability to observe molecular-level activities inside a living body through the injection of specific radioactive tracers. In addition to the commonly used F-18-FDG, new tracers are being constantly developed and investigated to pinpoint specific pathways in various diseases. New PET scanners are also being proposed by exploiting time of flight (TOF) information, enabling depth of interaction capability, and extending the solid angle coverage. To realize the full potential of the new PET tracers and scanners, there is an increasing need for the development of advanced image reconstruction methods. This grant application proposes a new framework for regularized image reconstruction that synergistically integrates deep learning and regularized image reconstruction. The new framework is enabled by the recent advances in machine learning, which provide a tool to digest vast amount information embedded in existing medical images. The proposed method embeds a pre-trained deep neural network in an iterative image reconstruction framework and uses the deep neural network to regularize PET image directly. By training the deep neural network with a large amount of high-quality low-noise PET images, the proposed method can capture complex prior information from existing inter-subject and intra-subject data and thus is expected to substantially outperform the current state-of-the-art regularized image reconstruction method. The two specific aims of this exploratory proposal are (1) to develop the theoretical framework to synergistically integrate deep learning in regularized image reconstruction for PET and (2) to implement the proposed method and validate its effectiveness using existing animal data. Once the proposed method is validated using existing animal data, we will seek funding to acquire necessary human data for the implementation of the proposed method on clinical PET scanners.