Biomedical Imaging and Biomechanics
Images are an important part of biomedical knowledge. New computational technologies provide the opportunity to supplement traditional two-dimensional biological and medical images with dynamic three-dimensional images that can be viewed, rotated, and reversibly dissected in a manner analogous to the physical objects they represent.
Visible Human Project
NIH/NLM is building and evaluating digital image libraries of human anatomy. Full use and understanding of the anatomical structures depicted in such libraries require the integration of HPCC technologies with technologies used in medical imaging systems including computed tomography (CT) and magnetic resonance imaging (MRI). Combining this library with efficient rendering algorithms will provide new educational tools for researchers, healthcare providers, students, and the general public. NIH/NLM is working with industry and academia to encourage the development of interoperable methods for representing and communicating such electronic images.
The Visible Human Data Set consists of images from a male and a female cadaver. The data set derived from the male cadaver was completed in November 1994 and made available both by Internet ftp (file transfer protocol) and DAT tape; NCAR assisted in distributing the data set. The size of the male data set is about 15 GB. More than 150 non-financial license agreements to access the data set have been signed by government, commercial, and academic organizations. Proposed applications include multimedia educational materials for educational levels from K through post- graduate medical education and home education, virtual reality programs includng surgical simulators, and modeling applications.
The data set from the female cadaver, which is expected to become available early in FY 1996, will be more detailed than the male.
In order to facilitate retrieval, methods need to be developed to link image data to symbolic text-based data that includes names, hierarchies, principles, and theories. Basic research is needed in the description and representation of structures and the connection of structural-anatomical to functional-physiological knowledge. The long-term project goal is to link the print library of functional- physiological knowledge with the image library of structural- anatomical knowledge transparently into one unified resource of health sciences information.
Cryosectional image from the Visible Human Male.
A PET image is formed through a computational reconstruction process. The computational time for the reconstruction and the quality of the resultant image depend primarily on the reconstruction algorithm that is used. Fourier methods have traditionally been used -- they are fast but can lead to artifacts. The iterative expectation maximization method (EM) based on a maximum likelihood criterion is known to yield reconstructions that are as good as or better than Fourier methods and have lower patient dose, but their computational demands have limited their use. New generation PET scanners allow for the retraction of their lead shields, and current research is studying whether the additional information that is gathered improves the quality of the image, though at the price of increased computation. The EM algorithm has been implemented on the Intel iPSC/860 at NIH/DCRT for use with the GE Advance scanner, but can be used for other three-dimensional scanners. One iteration of the EM algorithm for reconstructing a brain-size image using data from retracted scanners on 32 processors in the Intel system took 55 minutes. Reconstructions of body-size images will require larger and faster computing systems, more memory, faster I/O to disk, and improved algorithms.
Image Processing of Electron Micrographs
Large icosahedral viruses can be reconstructed using high-resolution electron microscopy and three-dimensional image reconstruction. The input to the reconstruction is a set of two-dimensional projections of virus particles obtained from electron micrographs. Estimating the two-dimensional orientation of these particles and using that information to perform the three-dimensional reconstruction are computationally intensive and memory-intensive calculations. Parallel algorithms have been developed to (1) estimate each particle's orientation by distributing the particles and associated computations across the computing systems' processors, and (2) reconstruct the complete three-dimensional image by distributing the micrograph data to balance the computations across the processors. The discovery of the location of individual proteins in the human herpes virus was due in part to the use of and improvements to these algorithms. This work was conducted at DCRT.
Three-dimensional reconstruction of large icosahedral viruses. Shown are images of herpes simplex virus type 1 capsids, which illustrate the potential of new parallel computing methods. They show the location of a minor capsid protein called VP26 as mapped in experiments in which VP26 was first extracted from purified capsids by treatment with guanidine hydrochloride and then rebound to the capsids. The right half of the top image shows the depleted capsid and the rebound VP26 capsid, and the left half shows the three-dimensional reconstruction, as it would be obtained with a conventional sequential computer. Parallel computing extended the analysis to obtain the lower images, which improved the signal-to- noise ratio and the resolution from approximately 3.5 to under 3.0 nanometers. The clusters of six VP26 subunits, shown together in the top image, are clearly resolved in the bottom image. This work was conducted at NIH in collaboration with the University of Virginia.
A key feature of light microscopy is the ability to observe movement and activity in living cells. In this NSF-funded Grand Challenge at the center for Light Microscope Imaging and Biotechnology at Carnegie Mellon University (an NSF Science and Technology center), these microscopes will be coupled with high performance computing systems that are up to 1,000 times more powerful than they were several years ago. This environment will make possible:
A related "Collaboratory for Microscopic Digital Anatomy" National Challenge is described in Section II.7.
Light microscopy images of a keratocyte (an epidermal cell of a fish scale) obtained using Differential Interference Contrast to enhance visibility of nearly transparent structures both in raw form (left) and with boundary outlined by the "snake" method (right). Beginning with a curved line in the general vicinity of the boundary, the method then improves the fit by compromising between visible edges and preference for smooth boundaries without breaks or kinks not clearly dictated by the image. Edge finding is needed in automated determination of cell types, analysis and recording of cell motion, and searching for unusual events. One application area is automated techniques in medical diagnosis.
Joints in the human musculoskeletal system (for example, the knee) carry large forces when functioning normally, and potentially damaging loads in extreme and traumatic events. These forces pass through relatively thin layers of soft hydrated tissue, articular cartilage, which must function as frictionless, load-bearing surfaces. The thin layers slide over each other in what an engineer would term a sliding contact problem.
The goal of this Grand Challenge is to understand this sliding contact problem, enabling better understanding of how a joint carries load in normal and pathological situations and contributing to improved clinical treatment. High performance computing resources are needed to conduct the precise three-dimensional simulations of contact of two-phase (consisting of solid and fluid parts) tissues over complicated three-dimensional layer shapes with realistic loading conditions and material properties.
Underlying the overall goals of this project are research studies of importance and interest to the scientific computing community. They include (1) computational methods to solve the sliding contact problems of three-dimensional bodies, (2) methods to automatically control the error in a large scale numerical simulation, (3) constructing three-dimensional solid computer models of anatomical entities that have been digitally measured by methods such as CT scan, MRI, stereophotogrametry, or three-dimensional digitization, (4) parallel algorithms that can be used to solve nonlinear time-dependent problems (such as the deformation of soft tissues, in this project), and (5) using computer predictions with experimental data to help understand complicated nonlinear materials such as soft tissue.
This is a five year interdisciplinary effort involving bioengineering, computational mechanics, mechanical engineering, mathematics, and computer science supported by NSF at Rensselaer Polytechnic Institute and Columbia University.