Computational imaging and intelligent specificity (Anastasio)

摘要 在这个技术研究和开发(TRD)项目中，高级计算和机器学习 将开发满足与图像形成和图像分析相关的各种需求的方法 高分辨率无标签光学显微镜。计算方法正在迅速部署，这些方法是 改变测量数据的获取方式，改进显微镜的形成和分析 图像。这种方法对无标记显微镜领域的潜在影响是非常高的，并且可以最优地 以创新和信息丰富的方式利用内在的内生对比机制。已开发的 方法将成为拟建中心中许多项目的使能技术。这项研究将是 了解并与研究与开发和推动生物项目共同开发和评估。总的主题 这项工作的一部分是将基于成像科学、物理和深度学习(DL)的方法整合到 绕过无标签成像的限制和使用客观图像质量测量来系统地 验证和提炼所开发的方法。本课程将研究三大类计算方法。 这将使(1)无标签图像的图像到图像映射能够提供计算特异性， 改进的语义分割和/或增强的空间分辨率；(2)改进的图像重建 3D细胞成像；以及(3)从多通道无标记图像中提取生物相关信息 数据。该项目的具体目标是：目标1：提供特异性的图像到图像的翻译方法， 语义分割和/或增强的空间分辨率；目标2：衍射层析成像和逆成像 用于3D成像的散射方法；以及目标3：发现生物标记物和多模式DL方法。 这个项目的成功完成将导致计算和动态链接库方法的发展 无标签成像技术。这些方法将改进计算染色，增强 多通道无标签图像的空间分辨率、语义分割、3D图像形成和分析 数据。它们将被系统地验证用于生物医学应用，这些应用属于 拟建的P41中心。所有源代码、经过培训的模型和文档都将开源并 在网上分享。

SUMMARY In this technology research and development (TRD) project, advanced computational and machine learning methods will be developed that address a variety of needs related to image formation and image analysis in high-resolution label-free optical microscopy. Computational methods are being rapidly deployed that are changing the way that measurement data are acquired and improving the formation and analysis of microscopy images. The potential impact of such methods on the field of label-free microscopy is very high and can optimally leverage inherent endogenous contrast mechanisms in innovative and informative ways. The developed methods will serve as enabling technologies for many projects in the proposed center. The research will be informed by and jointly developed and evaluated with the TRD and driving biological projects. A general theme of this work is the integration of imaging science, physics- and deep learning (DL)-based approaches to circumvent the limitations of label-free imaging and the use of objective image quality measures to systematically validate and refine the developed methods. Three broad classes of computational methods will be investigated that will enable the (1) image-to-image mapping of label-free images to provide computational specificity, improved semantic segmentation, and/or enhanced spatial resolution; (2) improved reconstruction of images for 3D cellular imaging; and (3) extraction of biologically relevant information from multi-modality label-free image data. The Specific Aims of the project are: Aim 1: Image-to-image translation methods for providing specificity, semantic segmentation, and/or enhanced spatial resolution; Aim 2: Diffraction tomography and inverse scattering methods for 3D imaging; and Aim 3: Biomarker discovery and multi-modal DL methods. This successful completion of this project will result in computational and DL methods that will advance a variety of label-free imaging technologies. These methods will enable improved computational staining, enhance of spatial resolution, semantic segmentation, 3D image formation, and analysis of multi-modality label-free image data. They will be systematically validated for use in the biomedical applications that are within the purview of the proposed P41 center. All source code, trained models and documentation will be made open-source and shared online.