| dcterms.abstract | 3D digital pathology provides extreme scale quantitative data with high potential for basic research in a wide scope, and becomes an emerging technology promising to support computer aided diagnosis. Quantitative analyses of 3D pathology images involve both deriving 3D micro-anatomic objects and their features, and exploring spatial relationships among a massive number of biological objects. However, this is challenged by the overwhelming data scale and 3D pathology complexity. Our goal is to create a scalable and effective 3D digital pathology analytics framework for large-scale 3D pathology imaging data. The framework will provide novel methods on pathology image registration, segmentation, and reconstruction to transform voxel-based information from extremely large-scale 3D imaging data to 3D spatial objects. It will also provide a highly effective and scalable 3D spatial data management and querying system that enables efficient discovery of spatial patterns of 3D pathology objects. For 3D pathology image analysis, we propose a dynamic multi-resolution approach for registration of slides from serial sections. We introduce an improved formulation with directed prior information on vessel tube probability within a variational level set framework for vessel segmentation. We propose a two-stage object cross-section association approach for 3D reconstruction with local bi-slide vessel mapping followed by a global vessel structure association via a geometrical model fitting method and Maximum A Posteriori (MAP) estimation. For 3D spatial queries and analytics, we develop iSPEED, a highly scalable and efficient in-memory based 3D spatial data management and querying system. To achieve low latency, iSPEED stores 3D objects in memory in a highly compressed form with successive levels of detail. To minimize search space and computation cost, iSPEED provides global spatial indexing in memory through partitioning at multiple levels. iSPEED provides an in-memory 3D spatial query engine, which can be invoked on-demand for running many instances in parallel. The parallelization of queries is implemented in, but not limited to, MapReduce. At run time, iSPEED dynamically decompresses only needed 3D objects at the specified level of detail, and creates necessary spatial indexes in-memory to accelerate query processing, such as on-demand inter-object indexing and structural indexing for individual complex structured objects like vessels. | |
