Advanced Image Analysis Tools for Super-Resolved MRI in Small Animals

摘要 动物模型成像在生物医学研究中发挥着关键作用，使基础研究成为可能 了解疾病过程，以及评估新疗法的转化研究。体内成像， 特别是，它提供了对动物伤害最小的好处和测量发育的机会， 纵向变化。磁共振成像 (MRI) 是应用最广泛的活体成像之一 模式，因为它对多种生物参数具有出色的敏感性，并且具有不同的灵活性 动物模型，例如啮齿动物、雪貂和非人类灵长类动物。 MRI 不仅用于推进 了解神经退行性疾病，以及衰老、癌症、成瘾和心血管疾病。 然而，尽管研究界希望模拟临床试验并进行高通量 研究表明，动物模型中 MRI 的自动分析明显落后于最先进的工具 可用于人体成像分析。一个主要原因是大多数动物 MRI 采集都是两个- 具有高面内分辨率但厚切片的维度，而最强大的图像分析工具也可以工作 最适合各向同性采集。结果，许多研究人员已经放弃进行分析 涉及费力的手工描绘。 我们的团队最近开发了一种新颖的算法，使用深度学习来提取各向同性的空间分辨率 来自标准各向异性 MRI 采集，无需外部高分辨率训练 数据，这些数据通常无法获得且难以获取。回顾性恢复各向同性空间的能力 与高分辨率相比，这些二维 MRI 采集的分辨率可以显着降低成本 分辨率各向同性采集。此外，它还通过启用更高级的分析开辟了可能性 关键图像处理算法，如配准和分割，将得到更准确的执行和 具有更高的自动化程度。因此，我们建议在此 R21 应用程序中实现以下具体目标：1) 优化和评估我们基于深度学习的无监督超分辨率动物 MRI 方法； 2） 开发和评估高维数据的超分辨率算法； 3）公开发布 开发的工具。我们的首要假设是，所提供的工具将显着提高敏感度 对生物标志物进行成像，从而提高统计能力并减少动物研究的规模和成本。的 所提出的分辨率增强与最先进的图像分析技术的结合也将 通过消除费力且可能不一致的手动描绘的需要来提高可重复性。 此外，这些努力将使临床前和临床试验能够以几乎相同的方式实施 分析管道。本申请是为了响应 PAR 19-369“动物的发展”而提交的。 用于研究的模型和相关生物材料。”

SUMMARY Imaging in animal models plays a key role in biomedical research, enabling both foundational studies for understanding disease processes, as well as translational studies evaluating novel therapies. In vivo imaging, in particular, offers the benefits of minimal harm to the animal and opportunities for measuring developmental, longitudinal changes. Magnetic resonance imaging (MRI) is one of the most extensively used in vivo imaging modalities because of its excellent sensitivity to a multitude of biological parameters and flexibility with different animal models, such as rodents, ferrets, and non-human primates. MRI has been used not only to advance understanding of neurodegenerative diseases, but also aging, cancer, addiction, and cardiovascular disorders. However, despite the research community’s desire for emulating clinical trials and performing high-throughput studies, automated analysis of MRI in animal models has significantly lagged state-of-the-art tools that are available in the analysis of human imaging. A major reason is that most animal MRI acquisitions are two- dimensional with high in-plane resolution but thick slices, whereas the most powerful image analysis tools work best on isotropic acquisitions. As a result, many researchers have been resigned to performing analyses involving laborious, manual delineations. Our team has recently developed a novel algorithm that uses deep learning to extract isotropic spatial resolution from a standard anisotropic MRI acquisition possessing without the need for external high-resolution training data, which is typically unavailable and difficult to procure. The ability to retrospectively recover isotropic spatial resolution from these two-dimensional MRI acquisitions allows for significantly reduced costs compared to high- resolution isotropic acquisitions. Moreover, it opens up the possibilities for more advanced analyses by enabling key image processing algorithms, such as registration and segmentation, to be more accurately performed and with greater automation. We therefore propose to perform the following Specific Aims in this R21 application: 1) Optimize and evaluate our deep learning-based unsupervised super-resolution approach for animal MRI; 2) Develop and evaluate a super-resolution algorithm for higher-dimensional data; 3) Publicly release the developed tools. Our overarching hypothesis is that the provided tools will enable significantly more sensitive imaging biomarkers, thereby increasing statistical power and reducing the size and cost of animal studies. The combination of the proposed resolution enhancement with state-of-the-art techniques for image analysis will also increase reproducibility by obviating the need for laborious, and potentially inconsistent manual delineations. Furthermore, these efforts will enable both pre-clinical and clinical trials to be implemented with nearly identical analysis pipelines. This application is being submitted in response to PAR 19-369, “Development of Animal Models and Related Biological Materials for Research.”