A Fast Deconvolution Algorithm for 3D Microscope Images

描述（由申请人提供）： 活细胞的时间推移3D显微镜成像广泛应用于许多生物医学研究领域。该技术可能在单个实验中生成数百个图像体积，这对于基本生物机制（如基因表达和蛋白质动力学）的可视化和量化至关重要。在进行进一步的图像分析和测量之前，通常对图像进行去卷积以提高对比度和分辨率。反卷积可能需要3-5个小时来处理15-30分钟实验中产生的数据，从而在数据流中产生瓶颈。目前的反卷积算法要么是鲁棒的，但对于大量的图像不够有效（例如MLEM算法），或快速，但对噪声敏感（例如黄金算法）。提出了一种新的高效和鲁棒的算法，该算法建设性地组装了当前领先的算法，以利用它们各自的优势。初步研究表明，它有可能通过几个因素提高反卷积速度。直接的好处是，反卷积时间显着减少，并增加了研究吞吐量。第一阶段的研究将集中在核心算法的开发上。样本数据集，主要是宽视场显微镜图像，将从合作者获得，可行性将根据公认的标准，通过检查去卷积图像中的特征结构进行评估。将对PSF在整个延时实验中如何变化的分析进行验证，以评估PSF是否需要在序列的去卷积过程中不断重新估计。第二阶段研究将进一步研究反卷积的加速方法，使用宽场、旋转盘共聚焦和激光扫描共聚焦显微镜图像彻底评估算法的性能，并将开发一个快速反卷积软件模块。第二阶段还将研究快速算法在处理非常大的单个数据量方面的应用。

DESCRIPTION (provided by applicant): The time lapse 3D microscope imaging of living cells is widely used in many biomedical research areas. This technique potentially generates hundreds of image volumes in a single experiment, which are critical to the visualization and quantification of fundamental biological mechanisms such as gene expression and protein dynamics. Deconvolution is often applied to the images to improve contrast and resolution before further image analysis and measurements are performed. Deconvolution may take 3-5 hours to process data produced in an experiment of 15-30 minutes, creating a bottleneck in the data flow. Current deconvolution algorithms are either robust but not efficient enough for the large number of images (e.g. MLEM algorithm), or fast but sensitive to noise (e.g. Gold's algorithm). A novel efficient and robust algorithm is proposed which assembles the current leading algorithms constructively to take advantage of their individual strengths. A preliminary study has shown its potential to increase the deconvolution speed by several factors. The immediate benefits are that the deconvolution time is significantly reduced, and the research throughput is increased. The Phase I research will focus on the development of the core algorithm. Sample data sets, primarily wide-field microscope images, will be obtained from collaborators, and feasibility will be evaluated by examining the feature structures in the deconvolved images based on commonly accepted criteria. An analysis of how the PSF varies throughout a time-lapse experiment will be undertaked to assess if the PSF needs to be continually re-estimated throughout the deconvolution of the sequence. The Phase II research will investigate further acceleration methods for deconvolution, thoroughly assess the performance of the algorithm using wide-field, spinning disk confocal and laser scanning confocal microscope images, and a fast deconvolution software module will be developed. Application of the fast algorithms to processing very large single data volumes will also be investigated in Phase II.