CAREER: Low Latency, Parallel, and Context Aware Vision in Computed Tomography

非侵入性体积成像技术的日益成熟和可用性为科学、工程和医学界提供了获取有生命和无生命物体的密集定量表示的强大方法-为我们提供了强大的信息捕获方法。这项研究工作集中在识别和标记捕获数据中的表示的方法，这些表示集成了先验的结构和空间配置知识。由体积扫描仪产生的体积像素(体素)的自动识别和标记提出了目前仍未解决的具有挑战性的不适定问题。即使对于训练有素的临床医生来说，在将解剖标签应用于低对比度医学图像时确定器官边界也可能是高度主观的，尽管他们拥有关于结构和形状的强大先前解剖学知识。因此，由不同的临床医生进行的解剖描述，即使是针对相同的患者图像，也不能提供一致性。更困难的不适定问题，比如在不同时间拍摄的患者图像之间的对应体素匹配--或者更困难的是，在两个完全不同的患者之间--是如此主观，以至于人类很少能提供一致的答案。这项研究的目的是为这些问题开发算法解决方案，以便提供跨卷的一致的定量分析；从而消除因无意失明或无意偏见而产生的主观性。随着进一步的进步，这种算法将足够快速和稳健，以提取解剖结构信息并在大规模自主执行患者通信；从而实现强大的数据分析，为数据驱动医学的未来铺平道路。通过课堂整合和课程开发，PI将培训理工科学生使用这种下一代图像处理算法。国际和平研究所将招聘和指导本科生和研究生两级的研究人员，重点是在与STEM相关的领域中招聘代表性不足的群体。因此，本研究与NSF推动科学进步、促进国民健康、繁荣和福祉的使命相一致。本研究项目包括三个主要工作：第一，基于空间关系先验的多结构同时分割(即位置感知分割)；第二，结构感知配准算法的开发(利用分割结果)；第三，针对数据并行计算机体系结构的这些算法的开发。PI开发的情境感知算法旨在执行同时的多目标分割，以及对器官变形和运动的解剖学专业推断，这些推理对成像不一致、设置变化和低剂量成像具有很强的鲁棒性。PI将研究对噪声、不完整或具有挑战性的图像采集具有稳健性的算法方法，方法是同时求解多个逆问题，并在一定的基础上结合解决方案和利用信号稀疏性获得更高的精度。这项研究产生的算法将对使用计算机断层扫描的领域产生广泛的社会影响，包括考古学、土壤科学、木材工业、生物科学、工业X射线检查和航空安全行业。除了传播新的算法外，PI还将开发高性能的并行实现，作为一个在允许的开源许可证下发布的库。开发将使用Git公开进行，从而实现敏捷、分散的开发，鼓励外部科学家和开发人员增加对项目的利用和贡献。最初选择纳入的算法设施是首席研究人员的专业领域，涵盖了广泛的应用，包括运动估计、图像拼接、分割、3D体积重建(计算机断层扫描)和配准/图像融合。通过联盟和讲习班，将鼓励领域专家在既有领域和新兴领域(如数字图像取证)中贡献他们的专业知识；促进跨领域领域的科学交流和协作。

The increasing availability and maturity of non-invasive volumetric imaging techniques has provided the scientific, engineering, and medical communities with powerful methods of acquiring dense quantitative representations of both animate and inanimate objects---providing us with powerful methods of information capture. This research effort concentrates on approaches for identifying and labeling representations within the captured data that integrate a priori structural and spatial configuration knowledge. The automatic identification and labeling of volumetric pixels (voxels) produced by volumetric scanners presents challenging ill-posed problems that presently remain unsolved. Even for well trained clinicians, determining organ boundaries when applying anatomical labels to low contrast medical images can be highly subjective, despite possessing strong prior anatomical knowledge of structure and shape. Consequently, an anatomical delineation performed by different clinicians, even for the same patient image, fails to provide consistency. More difficult ill-posed problems, such as matching corresponding voxels between images of a patient taken at different times---or, more difficult yet, between two entirely different patients---are so subjective that humans rarely provide consistent answers. This research effort aims to develop algorithmic solutions to these problems in order to provide consistent quantitative analysis across volumes; thereby removing subjectivity attributable to inattentional blindness or unintentional bias. With further advancement, such algorithms will be adequately fast and robust to extract anatomic structural information and perform patient correspondence autonomously at massive scales; thereby enabling powerful data analytics paving the way for the future of data driven medicine. Through classroom integration and curriculum development, the PI will train science and engineering students to work with this next generation of image processing algorithms. The PI will recruit and mentor researchers at both the undergraduate and graduate levels with emphasis on the recruitment of underrepresented groups within STEM related fields. Therefore, this research aligns with the NSF mission to promote the progress of science and to advance the national health, prosperity and welfare.This research project consists of three primary efforts: first, the simultaneous segmentation of multiple structures using spatial relationship priors (i.e. situationally aware segmentation); second, the development of structurally aware registration algorithms (leveraging segmentation results); and third, the development of these algorithms specifically targeted to data parallel computer architectures. Situationally aware algorithms developed by the PI will aim to perform simultaneous multi-target segmentations as well as anatomically specialized inference of organ deformation and motion that are robust to imaging inconsistencies, setup variations, and low-dose imaging. The PI will investigate algorithmic methods possessing robustness to noisy, incomplete, or otherwise challenging image acquisitions by solving multiple inverse problems simultaneously and achieving higher accuracy through the coupling of solutions and exploitation of signal sparsity under certain basises. The algorithms produced by this research will have broad societal impacts on fields employing computed tomography including archeology, soil sciences, the timber industry, biological sciences, industrial X-ray based inspection, and the aviation security industry. In addition to the dissemination of the novel algorithms, the PI will develop high-performance parallel implementations as a library released under a permissive open-source license. Development will occur in the open using Git; thereby enabling agile decentralized development that encourages increased utilization by and contributions from scientists and developers extramural to the project. Algorithmic facilities initially selected for inclusion are areas of principal investigator's expertise and cover a wide spectrum of applications including motion estimation, image stitching, segmentation, 3D volume reconstruction (computed tomography), and registration/image fusion. Through consortia and workshops, domain experts will be encouraged to contribute their expertise in established and emerging fields (e.g. digital image forensics); enabling scientific cross-fertilization and collaboration across domain specific fields.