Transcript An Engineering Research Center for Integrated Sensing and

Center for
Subsurface Sensing & Imaging Systems
Overview of Image
and Data Information
Management in
CenSSIS
David Kaeli
Northeastern University
Boston, MA
[email protected]
Overview of the Strategic Research Plan
Bio-Med
L3
Enviro-Civil
S2 S3
S1
S4
S5
Validating
L2 TestBEDs
R2
L1 Fundamental
Science
R3
R1
Image and Data
Information
Management
R3 Research Thrust Overview
 Utilize enabling hardware and software
technologies to address CenSSIS barriers
 Pursue research in enabling technologies
 Develop a common set of tools and techniques to
address SSI problems:
 Hardware parallelization and acceleration
 Software toolboxes
 Image database management and tools
Toolboxes
CenSSIS Middleware Tools
 Parallelization of MATLAB, C/C++ and Fortran codes
using Message Passing Interface (MPI) – a software
pathway to exploiting GRID-level resources
 Utilizing MPI-2 to address barriers in I/O performance
 Building on existing Grid Middleware such as Globus
Toolkit, MPICH-G2 and GridPort
 Presently illustrating the impact of the GRID on system
level projects (tomosynthesis reconstruction)
MPI
MATLAB
C/C++
Fortran
Parallelization
MPICH-G2
UPC
Air
Impact on CenSSIS Applications
Mine
Soil
 Reduced the runtime of a single-body Steepest Descent Fast
Multipole Method (SDFMM) application by 74% on a 32-node
Beowulf cluster
Scattered Light Simulation
• Hot-path parallelization
• Data restructuring
• Obtained superlinear speedup of Ellipsoid
Algorithm run on a 16-node IBM SP2
• Matlab-to-C compliation
• Hot-path parallelization
Runtime in seconds
• Matlab-to-C compliation
• Hot-path parallelization
100000
Original
10000
1000
Matlab-to-C
100
Hot path
parallelization
10
1
Ellipsoid Algorithm Speedup
(versus serial C version)
Speedup
 Reduced the runtime of a Monte Carlo
scattered light simulation by 98% on
a 16-node Silicon Graphics Origin 2000
Speedup
20
15
10
5
0
1
2
4
8
16
Number of Nodes
64-vector
1024-vector
256-vector
linear speedup
Tomographic mammography
 3D image reconstruction from x-ray projections
 Used to detect and diagnose breast cancer
 Based on well developed mammography techniques
 Exposes tissue structure using multiple projections from
different angles
 Advantages
 Accuracy: provides at least as much useful information
than x-ray film
 Flexibility: digital image manipulation, digital storage
 Structural information: using layered images
 Safe: low-dose x-ray
 Lower cost: compared to MRI
Image acquisition/reconstruction process
 Acquisition: 11 uniform angular
samples along Y-axis
 X-ray projection: breast tissue density
absorption radiograph
 Algorithm: constrained non-linear
convergence and iterative process
 Uses a Maximum Likelihood Estimation
Y
X-ray source
Initialization
Set 3D volume
Forward
Compute projections
3D volume
Backward
Correct 3D volume
Exit
Yes
Y
No
Satisfied ?
Z
X-ray
projections
X
detector
Parallelization approaches
 Reduce communication data
 Segmentation along Y-axis
 Using redundant computation to replace communication
 Segmenting along x-ray beam
First approach:
Non inter-communication
(more computation, less
communication)
Overlap area
Second approach:
Overlap with
inter-communication
exchange data
Third approach:
Non-overlap with
inter-communication
(less computation, more
communication)
Tomosynthesis Acceleration
•Input data set: phantom
1600x2034x45
Phantom data test results using nonoverlap method on 32 CPUs
• Serial implementations runs in 23 hours on a P4 machine
350
• Platforms:
– SGI Altix system
– UIUC NCSA Titan cluster
250
– P4 cluster at MGH
• Number of processors: 32
Computation: SGI Altix with
Itanium 2 processor outperforms
the other CPUs
 Currently moving this work to
the GRID and the Pittsburgh
Supercomputer Center
 Prototype running on our GRID
system at NU
Time (sec)
– UIUC NCSA IBM p690
– UMich Hypnos cluster
File IO
Collect
Inter-comm
Sync
Backward
Forward
300
200
150
100
50
0
P4 cluster
Hypnos
cluster
Titan
cluster
Platform
IBM p690
SGI Altix
Field Programmable Gate Arrays for
Subsurface Imaging
 Backprojection for Computed Tomography image
reconstruction
 Sponsored by Mercury Computer
 Finite Difference Time Domain (FDTD) in hardware
 Collaboration with Humanitarian Demining project
 Retinal Vascular Tracing in real time
 Collaboration with Real-time Retinal Imaging project
 Phase Unwrapping
 Collaboration with 3-D Fusion Microscope project
 Diverse problems, similar solutions:
FPGAs are particularly well suited for accelerating image processing
and image understanding algorithms
Retinal Vascular Tracing: Register 2-D Image to
3-D in Real Time
Objective
To accelerate an existing retinal
vascular tracing (RVT) algorithm by
implementing computation of template
responses in reconfigurable hardware
PCI
PCI BUS
BUS
RESULTS
MEMORY1
DESIGN
BLOCK RAM
Direction
of of
Direction
blood
vessel
blood
vessel
IMAGE
MEMORY0
FPGA
HOST
“Smart Camera”
FIREBIRD BOARD
Some Recent Publications on
Parallelization
• “Execution-Driven Simulation of Network Storage Systems,” Y. Wang and D.
Kaeli, Proceedings of the 12th ACM/IEEE International Symposium on
Modeling, Analysis of Computer and Telecommunication Systems, October
2004, pp. 604-611.
• “Profile-guided File Paritioning on Beowulf Clusters,” Y. Wang and D. Kaeli,
Journal o f Cluster Computing, Special Issue on Parallel I/O, to appear,
• “An Object-oriented Parallel Library,” C. Oaurrauri and D. Kaeli, International
Journal of High Performance of Computing and Networking, to appear.
• “Digital Tomosynthesis Mammography using a Parallelized Maximum
Likelihood Reconstruction Method,” T. Wu, R. Moore, E. Rafferty, D. Koppans,
J. Zhang, W. Meleis and D. Kaeli, Medical Imaging, 5368, February 2004.
• “Mapping and characterization of applications in Heterogeneous Distributed
Systems,” J. Yeckle and W. Rivera , To appear in Proceed. of the 7th World
Multiconference on Systemics, Cybernetics and Informatics (SCI2003).
• “Profile-Guided I/O Partitioning,” Y. Wang and D. Kaeli, Proceedings of the 17th
ACM International Symposium on Supercomputing, June 2003, pp. 252-260.
• “Source-Level Transformations to Apply I/O Data Partitioning,” Y. Wang and D.
Kaeli, Proceedings of the IEEE Workshop on Storage Network Architecture
and Parallel IO, Oct. 2003, pp. 12-21.
Held
again
in
2004
CenSSIS Solutionware – UPRM/NU/RPI
Toolbox Development
Toolboxes
 Support the development of CenSSIS Solutionware that
demonstrates our “Diverse Problems – Similar Solutions”
model
 Develop Toolboxes that support research and education
 Establish software development and testing standards for
CenSSIS
Image and Sensor Data Database
 Develop an web-accessible image database for CenSSIS
that enables efficient searching and querying of images,
metadata and image content
 Develop image feature tagging capabilities
Status of the CenSSIS Toolboxes
 Hyperspectral Image Analysis Toolbox
(HIAT)
HIAT
 October 2004
 Multiview Tomography Toolbox (MVT)
 fddlib:
 January 2003 (v. 1.0)
 July 2003 (v. 1.1)
MVT
 mvt:
 October 2004
 Rensselaer Generalized Registration
Library (RGRL)
RGRL
 September 2004
New toolbox: Improving the quality of
radiation oncology @ MGH
 Developed a 4D (3D + including time)
visualization browser tool kit
 Visualize Computed Tomography (CT)
images, organ outlines (wire contours) and
the isodose lines (treatment dosage)
 Present all this information in a user friendly
interface
4-D Visualization of Lung Tumors
Dosage
4-D Visualization
The Future for CenSSIS Toolboxes
SCIRun
Collaboration with the University of Utah
CenSSIS Image Database System
 Deliver an web-accessible database for
CenSSIS that enables efficient searching
and querying of images, sensor data,
metadata and image content
More that 4000 metadata-rich images/datasets
presently available online (> 10,000 by 2006)
 Database Characteristics:
• Relational complex queries (Oracle9i)
• Data security, reliability and layered user privileges
mouse
embryo
• Efficient search and query of image content and metadata
• Content-based image tagging using XML
• Indexing algorithms (2D, 3D, and 4D)
• Explore object relational technology to handle collections
3
4
2
1
CenSSIS Image Database System
CenSSIS Image Database System
CenSSIS Image Database System
Utilize Machine
Learning
algorithms to
improve query
view
CenSSIS Image Database System
Provides
data
description
associated
with initial
collection,
but does not
allow for
further
elaboration
or
annotation.
Image Annotation
 Provide the ability to markup image with searchable
features
 Enable image database to be more effectively datamined
<xml version=“1.0” encoding=“UTF-8”>
<embryo>
<description> Embryo developmental
stages</description>
<feature label=“1” xPos1=“29” yPos1=“33”
xPos2=“48” yPos2=“50”> 1 cell embryo </feature>
<feature label=“2” xPos1=“50” yPos1=“28”
xPos8=“70” yPos2=“40”> 2 cell embryo
</feature>
<feature label=“3” xPos1= “5” yPos1= “5”
xPos2=“25 yPos2=“20”> 4 cell embryo </feature>
</embryo>
XML and Java
• XML (Extensible Markup Language)
• Provides maximum flexibility and portability
• Well-supported standard
• Powerful querying tools available in Oracle
• The Java2 Platform
• Cross-platform compatibility
• Standard web-browser interface
• Native XML support
Image Tagging
A raw image file from the CenSSIS Database
• QUERY: I want to be able to add to this image textual
annotations, providing my medical team with questions about
particular ROIs:
• Difficult to describe regions in an image
• Difficult to pinpoint specific features in images
• Global image metadata too coarse to facilitate low-level tagging
Image Tagging
Image with tags
• Metadata associated with specific areas
• Query for specific image features
The Image Tagging Interface
Drawing
Tools
Tag
Options
View Options
Tags and XML
<feature type="Ellipse" label="4 Cell Stage">
<ellipse>
<xCenter> 101 </xCenter>
<yCenter> 58 </yCenter>
<xRadius> 79 </xRadius>
<yRadius> 46 </yRadius>
</ellipse>
<note> [custom XML tags go here] </note>
<annotator> awilliam </annotator>
</feature>
The Future Role of Image Annotation
 Provide a vehicle for natural collaboration
• A richer set of metadata to enable more detailed
queries
• Potential to perform extensive data mining on image
content
• An eye toward content-based image retrieval
 Tumor tracking paper recently accepted to SIGMOD 2005
The CenSSIS Image Database System
 Hosts the image and sensor data of CenSSIS (>500
images online)
 http://censsis-db1.ece.neu.edu/
 Provides metadata indexed image searching
 Uses XML tags to allow for easy information
interchange
 Evolved into a project-based management system,
allowing users to organize their data hierarchically
 Key issue: how do we develop collaboration tools
that increase the value of data stored in the
database?
 Presently exploring how best to integrate both
visualization and image annotation into the existing
framework (NIH proposal)
CenSSIS Image and Data Information
Management
 Addressing key research barriers in computational
efficiency, embedded computing and image/sensor data
management
 Exploiting Grid resources to enable new discovery in SSI
applications
 Producing a image/data repository and software-engineered
Subsurface Sensing and Imaging Toolsets
 Developing enabling tools targeting system-level projects
• Near real-time reconstruction and visualization
• Visualization of complex motion
• Predicting motion in image data using database indexing
techniques
MVT