APIII - Advancing Practice, Instruction & Innovation Through Informatics

A Parallel Image Registration Framework for Terabyte Sized Microscopy Datasets

Ashish Sharma, PhD (ashish@bmi.usc.edu); Kun Huang, PhD;Tony Pan, MS; Metin Gurcan, PhD; Joel Saltz, MD, PhD, Department of Biomedical Informatics, Ohio State University, Columbus, OH

Context: High resolution, 3D tissue reconstruction from digital microscopy images, is rapidly becoming a major component in pathology and certain domains of basic biological research.  Individual images are often tens of gigabytes in size and the size of an entire sectioned tissue can exceed a terabyte (TB) of data.  Processing and reconstructing from such large volume of data is beyond the computational capabilities of today’s computers.  While significant progress has been made in acquiring and viewing microscopy images at very high resolutions, the task of reconstructing the 3D structure from TB sized image datasets remains a computational challenge.  One of the critical pieces in 3D reconstruction is image registration (the task of calculating the relative orientation and deformation between two or more images).  Here we present a parallel image registration framework that partitions individual images across a cluster of computers, registers the partitioned image dataset and then aggregates the registered dataset for subsequent 3D reconstruction and explorative multimodal visualization.

Technology: Histological sections are prepared with standard hematoxylin and eosin (H&E) and immunohistochemical (IHC) protocols.  The sections are digitized using Aperio ScanScope slide digitizer.  The framework uses classification and registration algorithms developed in Matlab and C++.  The parallelization of these algorithms is achieved using the message passing libraries like MPI and MatlabMPI.  A cluster of computers connected to each other using high speed interconnects like Infiniband with TB’s of local storage space is needed to run this parallel framework.

Design:  The framework consists of an image partition component which can subdivide an image based on hardware parameters, feature size and other user specified domain parameters for subsequent processing and execution over a cluster of computers.  It is also responsible for aggregating the registered images for subsequent reconstruction, visualization and quantitative analysis.  The various registration and classification algorithms are largely treated as black boxes with well defined interfaces that provide information such as the transformations calculated by the algorithms.  As an additional optimization, a multiresolution image representation scheme is also provided which allows coarse but very fast rigid registrations to be performed on low resolution smaller images.  These fast registrations produce the initializations for subsequent fine grained registration algorithms which process images at full resolution.

Results: We have applied this framework to do some preliminary reconstruction of mammary ducts from an image dataset of serially sectioned mammary tissue.  Parallelized rigid and non-rigid deformable registration algorithms were used in this reconstruction.  A speedup of over 90% has been observed in these preliminary reconstructions.

Conclusions: This is a system that facilitates parallel execution of registration algorithms for 3D tissue reconstruction of massive image datasets.  It provides the functionality to modify the registration algorithm without having to deal with the complexities of image partitioning, aggregation and any intermediate data transfer. The parallel framework is also extensible to certain types of image segmentation and classification algorithms that employ short range local features.