CAREER: Probabilistic Framework for Self-Supervised, Data-Driven Computational Imaging

计算成像系统旨在改进传统成像设备，实现超分辨率或三维成像等先进功能。在计算机成像系统中，如计算机断层扫描（CT）、磁共振成像（MRI）和相位显微镜，对物体的间接测量是用专门的硬件进行的，然后用软件计算最终的重建。例如，CT中的测量结果是投影图像，软件后处理结果是物体的三维重建。对象是计算成像系统的输入，例如显微镜载玻片上的生物细胞或CT扫描中的大脑。计算成像系统的缺点是，它们通常需要对每个对象进行大量测量，这可能是昂贵的和缓慢的收集，使它们对许多应用程序望而却步。该程序开发数据驱动，概率方法的计算成像，以减少成像时间。与许多其他机器学习方法不同，开发的算法不需要基础事实或参考训练数据集。这个项目的成功将使生物学和医学的科学进步通过成像在以前无法进入的时空制度。更广泛的影响是广泛的，包括：降低CT和电子显微镜的剂量，缩短MRI的采集时间，减少荧光定位显微镜用于活细胞成像的光漂白，以及外科手术中的实时显微镜。已开发的方法可以应用于使医学成像更快，改善患者舒适度，减少辐射暴露和降低成本。补充教育计划将与研究计划同步发展，以扩大工程领域代表性不足群体的参与。项目包括为来自附近服务不足社区的高中生提供暑期体验，以及为本科生提供大学早期研究参与。提出的工作提高了发光二极管（LED）阵列显微镜的时间分辨率，这种模式允许在高分辨率和高视场的二维和三维重建定量振幅和相位（即介电常数）。LED阵列显微镜的算法开发将为其他计算成像方法（如微计算机断层扫描和x射线纳米全息断层扫描）的研究奠定基础。计算成像测量采用不同的硬件参数；在LED阵列显微镜的情况下，照明模式是不同的，以收集一堆图像。这项工作的关键创新之处在于联合重建相似的物体，每个物体只需要很少的测量次数，并且每个物体的硬件参数都不同。通过汇集来自整个对象集的测量信息并结合正演物理模型，可以联合推断先验和后验分布。为了有效地解决这个问题，这个程序通过变分自编码器的重新制定创造了一种新的技术。至关重要的是，没有假设地面真实数据集：只有对每个对象的噪声稀疏测量。与只产生点估计的重建算法相比，本工作中考虑的概率公式允许不确定性量化。模拟和实验数据将用于探索测量参数多样性对精度，稳定性和鲁棒性的影响。本文将研究一种新的自适应测量技术，该技术基于先前收集的目标数据选择用于目标测量的硬件参数。将创建和评估视频重建的高内存管理技术和考虑因素，其意义是允许将开发的框架应用于具有高度科学兴趣的实际实验问题。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Computational imaging systems aim to improve over conventional imaging devices, allowing for advanced capabilities such as super-resolution or three-dimensional imaging. In computational imaging systems, such as computed tomography (CT), magnetic resonance imaging (MRI), and phase microscopy, indirect measurements of an object are taken with specialized hardware, and software is used to compute a final reconstruction. For example, the measurements in CT are projection images, and software post-processing results in a 3-dimensional reconstruction of the object. The object is the input of the computational imaging system, such as biological cells on a microscope slide or a brain in a CT scan. The drawback of computational imaging systems is that they generally require a large number of measurements per object, which may be costly and slow to collect, making them prohibitive for many applications. This program develops data-driven, probabilistic methods for computational imaging to reduce imaging time. The developed algorithms will not require a ground truth or reference training dataset, unlike many other machine learning methods. Success of this program will allow scientific advances in biology and medicine by imaging in previously inaccessible spatiotemporal regimes. The broader impact is wide-ranging, including: lowered dose in CT and electron microscopy, reduced acquisition time in MRI, reduced photobleaching in fluorescence localization microscopy for live cell imaging, and real-time microscopy in surgical procedures. Developed methods can be applied to make medical imaging faster, improving patient comfort, reducing radiation exposure, and lowering costs. Complementary educational programs will be developed in tandem with the research plan to broaden participation from underrepresented groups in engineering. Initiatives include summer experiences for high school students from nearby underserved communities and early college research involvement for undergraduates. The proposed work improves the temporal resolution of light-emitting diode (LED) array microscopy, a modality that allows reconstruction of quantitative amplitude and phase (i.e., permittivity) in two and three dimensions with high resolution and high field-of-view. Algorithm development for LED array microscopy will build the foundation for research on other computational imaging methods such as micro-computed tomography and x-ray nano-holographic tomography. Measurements in computational imaging are taken with varying hardware parameters; in the case of LED array microscopy, the illumination pattern is varied to collect a stack of images. The key innovation in this work is to jointly reconstruct similar objects, each with a low number of measurements taken with hardware parameters that vary from object to object. By pooling information from measurements across the set of objects and incorporating the forward physics model, prior and posterior distributions can be jointly inferred. To efficiently solve this problem, this program creates a novel technique through a reformulation of variational autoencoders. Crucially, no ground truth dataset is assumed: only noisy, sparse measurements on each object. The probabilistic formulation considered in this work permits uncertainty quantification, in contrast to reconstruction algorithms that only yield a point estimate. Simulated and experimental data will be used to explore the impact of measurement parameter diversity on accuracy, stability, and robustness. A novel adaptive measurement technique will be evaluated, in which the hardware parameters for object measurement are chosen based on the object data previously collected. High memory management techniques and considerations for video reconstruction will be created and assessed, with the significance of allowing developed frameworks to be applied to real experimental problems of high scientific interest.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.