Improved Techniques for Substitute CT Generation from MRI datasets

该提案将改进替代CT图像，用于PET/MR和仅MR放射治疗 规划鉴于MR相对于CT的软组织对比度大大提高，这有助于PET的解释 对于PET/MR和放射治疗计划的靶勾画，剩余的限制是当前 仅从MR数据获得足够准确的替代CT图像的能力。不幸的是，MRI 分辨骨的能力有限，大多数MR采集无法区分空气和骨 使得这些组织类型的分割具有挑战性。该项目将利用深度学习，一个新的和 机器学习的发展领域，开发新的方法，从快速MR创建替代CT图像 可以在PET/MR和放射治疗计划工作流程中使用的采集。在目标1中，我们将研究 快速MR采集与深度学习方法一起用于头部和骨盆中的sCT生成 使用与PET/CT成像匹配的3 T PET/MR图像来创建深度学习训练和评估 数据集。不同的深度学习网络和MR输入将被研究和调整，以确定最佳的 PET重建性能。在目标2中，我们将研究快速但运动弹性的方法， 身体MR成像，用于后续基于深度学习的替代CT生成。在一个探索性的子目标中，我们 还建议研究仅利用仅PET数据的sCT生成方法。目标2中获得的数据 将用于创建全面的全身，运动弹性数据集，用于训练和评估深度 学习网络。在目标3中，我们将评估仅MR放射治疗的替代CT方法 规划将仅MR方法与基于CT的标准治疗模拟进行比较， 和颈部，胸部，腹部和骨盆，深度学习网络将针对区域进行优化和评估- 具体的RT计划和模拟。此外，还将研究迁移学习方法，以将sCT扩展到 0.35T MR-Linac，用于演示呼吸运动分辨替代CT生成。

This proposal will enable improved substitute CT images for use in PET/MR and MR-only radiation treatment planning. Given the greatly improved soft-tissue contrast of MR relative to CT, which aids interpretation of PET for PET/MR and target delineation for radiation treatment planning, a remaining limitation is the current capability to obtain sufficiently accurate substitute CT images from only MR-data. Unfortunately, MRI has limited capability to resolve bone and the inability of most MR acquisitions to distinguish between air and bone makes segmentation of these tissues types challenging. This project will utilize deep learning, a new and growing area of machine learning, to develop new methodology to create substitute CT images from rapid MR acquisitions that can be utilized in PET/MR and radiation treatment planning workflows. In Aim 1 we will study rapid MR acquisitions to be used with deep learning approaches for sCT generation in the head and pelvis using 3T PET/MR images matched with PET/CT imaging to create deep learning training and evaluation datasets. Different deep learning networks and MR inputs will be studied and adapted to determine the best PET reconstruction performance. In Aim 2 we will investigate rapid but motion-resilient approaches to whole- body MR imaging for subsequent deep learning-based substitute CT generation. In an exploratory subaim, we also propose to study methods of sCT generation that only utilize PET-only data. The data acquired in Aim 2 will be used to create comprehensive whole-body, motion-resilient datasets for training and evaluation of deep learning networks. In Aim 3 we will evaluate substitute CT approaches for MR-only radiation treatment planning. MR-only approaches will be compared to standard CT-based treatment simulation in the brain, head & neck, chest, abdomen, and pelvis and deep learning networks will be optimized and evaluated for region- specific RT planning and simulation. Additionally, transfer learning approaches will be studied to extend sCT to a 0.35T MR-Linac to demonstrate respiratory motion resolved substitute CT generation.