Medical data

Download Report

Transcript Medical data

WP10 - Biology
Medical Imaging on the grid
Johan Montagnat, WP10, DataGrid France, Thursday September 26th 2001
Medical Imaging on the grid
1. Medical applications for the grid
2. Grid specific requirements
3. Current developments
Johan Montagnat, WP10, DataGrid France, Thursday September 26th 2001
Medical data: the facts
• Generalization of the image for medical diagnosis and prognosis
• Digital imagers availability
• Software dedicated to medical image processing: on imagers / for
postprocessing
 medical image visualization
 quantitative processing
 remote services to health institutes
Huge amount of data produced by high resolution imagers
 High resolution image: 10243 voxels * 16 bits = 2 Gbytes / volume
 CHU Lyon: about 12 Tbytes of data per year on heart imagers
Johan Montagnat, WP10, DataGrid France, Thursday September 26th 2001
Grid added value for medical imaging
• Sharing computation resources
• Easy access to distributed databases
• Access to larger computation power / memory resources
Johan Montagnat, WP10, DataGrid France, Thursday September 26th 2001
Application 1: queries and computations on medical
databases
Objective:remote data access and management
Image
acquisition
and storage
center
Image
processing
center
physician
Distributed database access and visualization
Remote processing
Fast (almost interactive) response time for clinical usage
Johan Montagnat, WP10, DataGrid France, Thursday September 26th 2001
Application 2: Drug assessment
Objective: Remote processing of large image sets
Example: tagged MR images processing for cardiac activity quantification
tagged
MRI
sequence
1. Tags and myocardium
automatic extraction
time
2. Motion
estimation
3. Quantification
Johan Montagnat, WP10, DataGrid France, Thursday September 26th 2001
• Compute drug effect using large (remote) databases showing treated
patients and placebos.
 Chain a set of processings, pipe outputs into inputs
 Submit large datasets to the processing centers (large scale
parallelism)
 Enlarge datasets as more data become available
• As much as thousands of images involved.
Johan Montagnat, WP10, DataGrid France, Thursday September 26th 2001
Application 3: Interactive segmentation
One model
deformation step
Model visualisation
update
User correction
(interaction)
loop
Johan Montagnat, WP10, DataGrid France, Thursday September 26th 2001
Johan Montagnat, WP10, DataGrid France, Thursday September 26th 2001
Application 4: Physiology modeling
• Objectives: modeling organs anatomy, dynamics and physiology
for image analysis
bio-mecanical model (FEM)
electrical model
very complex structure
biological scale out of range
• Heart model-based segmentation
30 000 nodes model, ~ 1 Gbyte of memory
30 minutes of computation time / volume
Bioengineering research group, Auckland
• Heart model-based motion estimation
105 to 106 nodes model, more than 10 Gbytes of memory
FEM parallelization
Johan Montagnat, WP10, DataGrid France, Thursday September 26th 2001
Application 5: Surgery simulation
Objective: real-time model interaction
Biomedical
model
computation
Visualization
and force feedback
INRIA - Epidaure
Position tracking
Biomedical model-deformation
Real time visual (25 Hz) and force (300 Hz) feedback
Johan Montagnat, WP10, DataGrid France, Thursday September 26th 2001
Medical Imaging on the grid
1. Medical applications for the grid
2. Grid specific requirements
3. Current developments
Johan Montagnat, WP10, DataGrid France, Thursday September 26th 2001
Remote data storage and retrieval
• Medical files management inside/between health care centers
retrieval of medical information
 distributed database management, image indexing
• Networking with medical data
exchange between several institutes involved in the care of the same patient
 data privacy, access right control
Johan Montagnat, WP10, DataGrid France, Thursday September 26th 2001
Remote data processing
• Supervised algorithms
data visualization at interactive rate  interactive tools, compression, high
bandwidth networks
• Costly algorithms
 computation power and large memory requirements
• Data exchange format
 adaptability to algorithms input
• Algorithms management
 knowledge of existing processing services
• Priority levels
for surgery rooms, emergency situations…
 priority queuing, interruption of low priority jobs
Johan Montagnat, WP10, DataGrid France, Thursday September 26th 2001
Pipeline processing
• Medical trial of new drugs / large scale pathology studies
very large scale studies (100’s to 1000’s of images)
statistical interpretation of results
Input images
Noise filter
• Several processing stages
dataflow control
processing control
 pipeline architecture
Differential
operator
• Pipeline issues
Deformable models
one pipeline for multiple inputs
load balancing / synchronization
failure / retrial paradigm, logging
dynamic extension of the processed image set
Image labelling
Output images
Johan Montagnat, WP10, DataGrid France, Thursday September 26th 2001
Modeling and simulation
• New trend in medical image processing
building anatomical and physiological models to add a priori knowledge in
under-constrained direct and reverse problems
• Anatomical modeling and numerical simulation
Geometry and physiology modeling
Large scale discretization, modeling at biological scale
 Parallelizable numerical methods, fast message passing interface, large
memory
• Intervention simulation
Complex models
Real time visual and force-feedback capabilities
Johan Montagnat, WP10, DataGrid France, Thursday September 26th 2001
Medical Imaging on the grid
1. Medical applications for the grid
2. Grid specific requirements
3. Current developments
Johan Montagnat, WP10, DataGrid France, Thursday September 26th 2001
Remote data query / remote computation
Johan Montagnat, WP10, DataGrid France, Thursday September 26th 2001
MR Image simulator parallelization
• Simulator structure
Virtual
object
Magnetisation
computation
kernel
Reconstruction
algorithm
MRI
Image
MRI
sequence
• 3 level Parallelisation
source object components (isochromats) parallelization
kernel parallelization
sequence parallelization
• MPI implementation
Johan Montagnat, WP10, DataGrid France, Thursday September 26th 2001
Remote visualization - interactive processing
User workstation
Grid computing element
Graphic daemon
instantiate
User interface
process
connect
Interactive
application
communicate
Processing
loop
Johan Montagnat, WP10, DataGrid France, Thursday September 26th 2001