Signal Processing Laboratory Swiss Federal Institute of Technology, Lausanne 1 Supercomputers and client-server environment for biomedical image processing.

Signal Processing Laboratory Swiss Federal Institute of Technology, Lausanne

1

SupercomputersSupercomputersandand

client-server client-server environmentenvironment for biomedical image processing for biomedical image processing

C. Dufour and J.-Ph. Thiran

Signal Processing Laboratory Swiss Federal Institute of Technology, Lausanne

2In other words...

How to take advantage ofHow to take advantage ofsupercomputers processing powersupercomputers processing power

throughthroughclient-server applicationsclient-server applications

for biomedical image processingfor biomedical image processing

EPFL’s EPFL’s Swiss-TXSwiss-TX and and JavaJava applets applets

Signal Processing Laboratory Swiss Federal Institute of Technology, Lausanne

3Contents

Biomedical image processing factsBiomedical image processing facts

Supercomputers in a client-server environment.Supercomputers in a client-server environment.

EPFL’s supercomputer: the EPFL’s supercomputer: the Swiss-TXSwiss-TX

JavaJava based client-server applicationsbased client-server applications

Fundamentals on image processing parallellizationFundamentals on image processing parallellization

A practical example: the A practical example: the MicrotubulesMicrotubules application application

ConclusionsConclusions

Signal Processing Laboratory Swiss Federal Institute of Technology, Lausanne

4Biomedical image processing facts

Biomedical image processing is Biomedical image processing is very demanding in very demanding in

computer power computer power because of…because of…– Image data size

Electron microscopy 512x512 images 2 Mb X-Ray 2048x2048 images 32 Mb MRI 256x256x100 images 50 Mb Confocal microscopy 512x512x50x3 images 300 Mb

– Processing algorithm complexity Many processing operations required to reach the final solution Some operations may be highly iterative Some algorithm may be very complex

Signal Processing Laboratory Swiss Federal Institute of Technology, Lausanne

5Biomedical image processing facts

Examples of such applications are…Examples of such applications are…– 3D image registration

MRI-CT rigid registration Brain atlas matching (non-rigid registration)

– 2D image processing Software tool for early diagnosis of malignant melanoma

– 3D object based image compression Heart image compression

– 2D and 3D microscopy image analysis The Microtubules image analysis application Correlation analysis of pre- and post-synaptic proteins (from Confocal

images)

Signal Processing Laboratory Swiss Federal Institute of Technology, Lausanne

6Biomedical image processing facts

Processing time for large images and/or complex Processing time for large images and/or complex

algorithms may be algorithms may be particularly highparticularly high..

Processing time may be Processing time may be reduced significantly reduced significantly

through use of supercomputers and proper through use of supercomputers and proper

parallellization of processing algorithms.parallellization of processing algorithms.

Signal Processing Laboratory Swiss Federal Institute of Technology, Lausanne

7

Network

Image data

Supercomputers in a client-server environment

Allow Allow anyany useruser (biologist, physician) to take (biologist, physician) to take advantage of supercomputer(s) processing power advantage of supercomputer(s) processing power through the networkthrough the network

PC Supercomputer

AcquisitionViewing

Processing

Se

rverC

lien

tC

lien

t

Signal Processing Laboratory Swiss Federal Institute of Technology, Lausanne

8EPFL’s supercomputer: the Swiss-TX

The The Swiss-TXSwiss-TX is EPFL’s new supercomputer. is EPFL’s new supercomputer.Its main characteristics are :Its main characteristics are :

Signal Processing Laboratory Swiss Federal Institute of Technology, Lausanne

9EPFL’s supercomputer: the Swiss-TX

The The Swiss-TXSwiss-TX is EPFL’s new supercomputer. is EPFL’s new supercomputer.Its main characteristics are :Its main characteristics are :– Commodity based supercomputer

Signal Processing Laboratory Swiss Federal Institute of Technology, Lausanne

10EPFL’s supercomputer: the Swiss-TX

The The Swiss-TXSwiss-TX is EPFL’s new supercomputer. is EPFL’s new supercomputer.Its main characteristics are :Its main characteristics are :– Commodity based supercomputer– Up to 504 processing nodes (Swiss-T2)

Machine Setting up Processors Peak performance

[Gflops/s]

Memory

[GBytes]

Operating System

Swiss-T0 Dec. 97 8 8 2 Digital Unix

Swiss-T0 Dual Oct. 98 16 16 8 Windows NT

Swiss-T1 Apr. 99 72 72 36 Digital Unix

Swiss-T2 1stQ. 00 504 504 252 Digital Unix

Signal Processing Laboratory Swiss Federal Institute of Technology, Lausanne

11Java based client-server applications

JavaJava new programming language is particularly well new programming language is particularly well suited for network based suited for network based client-serverclient-server applications applications

JNI

RMI

Machine independentstand-alone application

Web browserapplets

Remote processingof image data

Efficient implementation of processing algorithm

Signal Processing Laboratory Swiss Federal Institute of Technology, Lausanne

12Fundamentals of image processing parallellization

Image processing algorithm may be easily Image processing algorithm may be easily parallelizedparallelized, either…, either…– splitting the image in tiles, each being processed on a

different node

Node 1

Node 3 Node 4

Node 2Overlap

Signal Processing Laboratory Swiss Federal Institute of Technology, Lausanne

13Fundamentals of image processing parallellization

Image processing algorithm may be easily Image processing algorithm may be easily parallelizedparallelized, either…, either…– splitting the image in tiles, each being processed on a

different node– splitting an algorithm in independent tasks, each node

taking care of its own task (though not all algorithm may be split in such a way)

Filter

Node 1Filter

Filter Node 2

Filter Node 3

Filter Node 4

Signal Processing Laboratory Swiss Federal Institute of Technology, Lausanne

14A practical example: the Microtubules application

Electron microscopy images analysis applicationElectron microscopy images analysis application((LTS-IBCM partnershipLTS-IBCM partnership))Automated statistical analysis of proteins densities, related

to microtubules proximity

Proteins

Microtubules

MICROTUBULES ANALYSIS - (c)1998 LTS (C.Dufour)

Processing... DONEComputing statistics... DONE Image surface (pixels) : 345600 Microtubules surface (pixels / %) : 97772 / 28.29 Microtubules total length (pixels) : 10347 Microtubules avg. width (pixels) : 9.45 Markers (0) quantity (elmts) : 197 Markers (0) near microtubules (elmts / %) : 126 / 63.00 Markers (0) avg. size (pixels) : 9.61 Markers (1) quantity (elmts) : 0 Markers (1) near microtubules (elmts / %) : 0 / 0.00 Markers (1) avg. size (pixels) : 0.00

Signal Processing Laboratory Swiss Federal Institute of Technology, Lausanne

15A practical example: the Microtubules application

The The image processingimage processing aspect aspectExtract binary masks representing the interesting

structures in the image (segmentation problem)

Signal Processing Laboratory Swiss Federal Institute of Technology, Lausanne

16A practical example: the Microtubules application

Swiss-T0Swiss-T0 processing timesprocessing times

0

0.25

0.5

0.75

1

1.25

C.1 C.2 C.4 C.6

Nor

mal

ized

pro

cess

ing

time

Minimum Average Maximum

0123456789

10111213

C.1 C.2 C.4 C.6

Tim

e ga

in f

acto

r

Best theoretical Minimum

Average Maximum

Signal Processing Laboratory Swiss Federal Institute of Technology, Lausanne

17Conclusions

Biomedical image processing is very demanding on computer Biomedical image processing is very demanding on computer power.power.

Use of supercomputers allows to Use of supercomputers allows to reduce processing timereduce processing time significantly.significantly.

Network based client-server applications allow to take Network based client-server applications allow to take advantage of supercomputer(s) processing power, advantage of supercomputer(s) processing power, easilyeasily and and at at lowest costlowest cost..

EPFLEPFL Swiss-TX Swiss-TX andand Web browser basedWeb browser based JavaJava applets are an applets are an elegant solution for all EPFL and LTS partners. elegant solution for all EPFL and LTS partners.

Early ‘99, a new Early ‘99, a new CTI project CTI project will bring LTS, STX Corp., UNIL, will bring LTS, STX Corp., UNIL, CHUC and HUG to collaborate in this new framework.CHUC and HUG to collaborate in this new framework.

Signal Processing Laboratory Swiss Federal Institute of Technology, Lausanne

18

Thank you for your attention !Thank you for your attention !