The architecture and performance of CAVASS Computer Assisted Visualization and Analysis Software Sys

1
The architecture and performance of CAVASS
(Computer Assisted Visualization and Analysis
Software System)
George J. Grevera, Jayaram K. Udupa, Dewey
Odhner, Ying Zhuge, and Andre Souza
Medical Image Processing Group Department of
Radiology - University of Pennsylvania Philadelph
ia, PA Department of Mathematics and Computer
Science Saint Josephs University Philadelphia,
PA
2
Introduction
Purpose To present the architecture and
performance of a new cluster-based open-source
software system called CAVASS (next incarnation
of 3DVIEWNIX). Goal of CAVASS To achieve
practical processing (image analysis and
visualization) time on even very large data sets.
CAVA Computer-Aided Visualization and
Analysis CAVASS CAVA Software System CAVA
deals with the science underlying computerized
methods of image processing, analysis, and
visualization to facilitate new therapeutic
strategies, basic clinical research, education,
and training.
3
CAVA Operations in CAVASS
Image processing for enhancing information
about and defining object system in
images. Visualization for viewing and
comprehending object system in its full form,
shape, and dynamics. Manipulation for altering
object system (virtual surgery). Analysis for
quantifying information about object system.
CAVA operations take object system information
from one space to another typically, and
eventually also to a quantitative space.
4
Previous software systems brought out by our
group
DISPLAY mini computer frame buffer
1980 DISPLAY82 mini
computer frame buffer 1982
(distributed to gt 150
sites with source.) 3D83 GE CT/T
8800 1983 3D98 GE CT/T
9800 1986 3DPC
PC-based 1989 3DVIEWNIX
Unix, X-Windows 1993
(distributed with source to 100s
of sites.) CAVASS platform
independent, wxWidgets 2008
5
CAVASS Target User Groups UG1 CAVA basic
researchers/technology developers UG2 CAVA
application developers UG3 Users of CAVA
methods in clinical research. Current Open
Source Software Limitations that CAVASS Attempts
to Address LM1 Lack of comprehensiveness of
CAVA operations. LM2 Lack of coverage for user
groups UG1-UG3. LM3 Lack of adequate
speed/efficiency. LM4 Lack of adequate
interfaces.
6
Methods
Key Features of CAVASS (F1) Open-source, C/C,
wxWidgets. (F2) Inherits most CAVA functions of
3DVIEWNIX. (F3) Incorporates most commonly used
CAVA operations, but does not go overboard on
generality and inclusiveness. (F4) Optimized
implementations for efficiency. (F5) Time
intensive operations parallelized and implemented
using Open MPI on a cluster of workstations
(COWs). (F6) Interfaces to popular toolkits (ITK,
VTK), CAD/CAM formats, DICOM support, other
popular formats. (F7) Stereo interface for
visualization.
7
Architecture

• Portable, standardized user interface (C,
  wxWidgets)
• Parallelization (MPI)
• Software engineering practices (doxygen, CMake,
  CVS)
• Interface to ITK
• Support for a wide variety of data formats
  (DICOM, GIF, JPEG, PNM, STL, TIFF, VTK)

8
Parallelization of CAVA Operations in CAVASS
Divide the input image into chunks and assign
each chunk to a processor. A chunk represents
data contained in a contiguous set of slices,
either image or object structure data.
CAVA operations can be divided into the following
three groups. Type 1 Operation
chunk-by-chunk, each chunk accessed only
once. Ex slice interpolation. Type 2 As
in Type 1, but significant further operation
needed to combine results. Ex 3D
rendering. Type 3 Operation chunk-by-chunk, but
each chunk may have to be
accessed more than once. Ex graph traversal.
CAVASS parallelizes all three groups of
operations.
9
Results
Test Data Sets Sequential and parallel
implementations of several Type 1 and Type 3
operations in CAVASS and ITK/VTK are compared
using three data sets Regular 256?256?46
MR brain image 6 MB Large 512?512?459
CT of thorax 241
MB Super 1023?1023?417 CT of head
873 MB (visible woman)
Platforms Multiprocessor system 3.4 GHz, dual
processor, 4GB RAM. COW 3.4 GHz single processor
systems, 1GB/sec connection.
10
In the following tables, the number of processors
used is shown in square brackets under
parallel. The times reported are in seconds. No
entries indicate that the operation was either
not tested or not available.
11
(No Transcript)
12
VTK rendering is assisted by a graphics
processor. CAVASS rendering is entirely in
software.
13
Conclusions

• COWs are more cost/speed effective than
  multi-processing systems. They are seemlessly
  expandable and upgradeable without requiring
  software changes.
• Most CAVA operations can be accomplished in
  reasonable time even for extremely large data
  sets on COWs in portable software.
• (3) COWs can be built quite inexpensively in CAVA
  research labs with publicly available hardware
  and software and standards.
• (4) CAVASS can handle extremely large data sets.
  It seems to be considerably faster than ITK in
  many image processing operations.

Further Information www.mipg.upenn.edu/cavass
CAVASS release date June 2008.
14
CAVASS example module
15
CAVASS DICOM explorer
16
Example of CAVASS integration with ITK
17
T-shell surface rendering in CAVASS