TR&D2: Advanced Statistical Image Reconstruction & Physics Informed Artificial Intelligence for Quantitative PET/MR

人工智能(AI)方法论的发展具有深远的重要性和预期 产生重大的社会影响，特别是对医学的影响。在过去的融资周期中，我们开创了 深度神经网络(DNN)在各种图像重建任务中的应用 对其在医学成像中的应用有广泛的了解和经验。在这个新的研发中，我们 提出用深度学习来推动人工智能在医学影像中的应用超越传统 图像重建问题。新的对比剂机制的研究(例如，新的MRI序列和 新的PET示踪剂)是PET/MR创新的主要前沿。为了改善图像质量，我们 将解剖图像和运动校正结合到一种新的基于DL的图像重建中 框架。我们还将基于MGH积累的大量成像数据构建我们的AI模型，同时 还提供了一种方法，将这些知识转移到现有数据很少的新研究中。我们的 提出的领域自适应和领域自适应少镜头学习技术将主要解决 人工智能在医学成像领域面临的一些最大挑战，即有限的训练数据、通用性 问题，从而使人工智能在临床环境中得到实际传播和使用。最后，我们 提出了重建图像的后验分布估计方法。可供使用的 重建的不确定性将为更优雅和准确的诊断打开新的窗口 精确医学中的方案和早期治疗反应评估因此导致显著的 PET/MR的新应用数量

The development of artificial intelligence (AI) methodology is of profound importance and is expected to have major societal impact, especially its effect on medicine. In the past funding cycle, we pioneered the application of deep neural networks (DNN) in various image reconstruction tasks and built a solid understanding and extensive experience of its applications in medical imaging. In this new TR&D, we propose use deep learning (DL) to push the application of AI in medical imaging beyond the traditional image reconstruction problem. The study of novel contrast mechanisms (e.g. new MRI sequence and new PET tracer) is a major frontline of PET/MR innovation. To achieve improved image quality, we will incorporate anatomic image and motion correction in a novel DL-based image reconstruction framework. We will also build our AI model based on the accumulated big imaging data at MGH while also providing a methodology to transfer this knowledge to new studies with few existing data. Our proposed domain adaptation and domain adaptation few shot learning technology will largely address some of the biggest challenges of AI in medical imaging, i.e., limited training data, the generalizability problem, thus enabling AI to be practically disseminated and used in clinical environment. Finally, we propose to estimate the posterior distribution of the reconstructed image. The availability of the uncertainty of reconstruction will open a new window for much more elegant and accurate diagnostic protocols and early treatment response evaluation in precision medicine thus leading to a significant number of new applications for PET/MR.