Date post: | 11-Jan-2016 |
Category: |
Documents |
Upload: | kathleen-page |
View: | 213 times |
Download: | 0 times |
CC Workshop May, 27th, 2004
Piergiorgio Cerello ([email protected]) 1
MAGIC-5Medical Applications on a
Grid Infrastructure Connection
ComputerAssistedDiagnosis (CAD)
Distributed ComputingInfrastructure (GRID)
&
INFN: Bari, Cagliari, Catania, Lecce, Napoli, Pisa, TorinoUniversities: Bari, Genova, Lecce, Napoli, Palermo,
Piemonte Orientale, Pisa, SassariHospitals: Alessandria, Bari, Livorno, Milano, Napoli,
Palermo, Pisa, Sassari, Torino, Udine
HEP expertise on Image Analysis (CAD) - CALMA Grid Computing
International Collaborations(CERN, CEADEN)
Agreement with BRACCO
CC Workshop May, 27th, 2004
Piergiorgio Cerello ([email protected]) 2
CALMA Breast Cancer Screening
Increased survival rate
Problems: costs and manpower
Computer Assisted Detection
Specificity (negatives/true negatives)
Sensitivity (positives/true positives) 73% - 88%
83% - 92%
2% - 10% increase with double reading
CC Workshop May, 27th, 2004
Piergiorgio Cerello ([email protected]) 3
CALMA Results Largest Database of digitised mammograms ( > 5000) ROC (Receiver Operating Characteristic) Curve
Massive Lesions
SENSIBILITY: 92%SPECIFICITY: 92%
Microcalcifications
SENSIBILITY: 94%SPECIFICITY: 95%
87.1 (4.0)82.9 (4.5)71.5 (5.4)C
90.0 (3.6)88.2 (3.8)80.0 (4.8)B
94.3 (2.8)94.3 (2.8)82.8 (4.5)A
+ CALMA+ CAD XRadiologist
70.9 (4.1)70.8 (4.2)74.2 (4.0)C
88.4 (2.9)85.9 (3.2)91.7 (2.6.)B
87.5 (3.0)84.2 (3.3)87.5 (3.0)A
+ CALMA+ CAD XRadiologist
Improved Sensitivity & Reduced Specificity
CC Workshop May, 27th, 2004
Piergiorgio Cerello ([email protected]) 4
2001: CALMA Open Issues Virtually unlimited Database size
Intrinsically distributed Database – many sources
Network connections
Access required to all the images
The “GRID philosophy” in mammographic CAD
Example: Italy4 mammograms/exam (60 MB)/exam6.7 Mpeople, 1 exam/2y 3.35 Mexams/year about 200 TB/year
1 PB/year on the European scale Huge amount of distributed data
Use CasesLarge Scale Screening
Teleradiology: diagnosis & training
CAD on demand
CC Workshop May, 27th, 2004
Piergiorgio Cerello ([email protected]) 5
UserInterface
InformationSystem
ComputingElement
Data & MetadataCatalogues
StorageElement
AuthenticationAuthorisation
MonitoringAccounting
“Green” VO
“Blue” VO
WM
S
CC Workshop May, 27th, 2004
Piergiorgio Cerello ([email protected]) 6
Medical Imaging communities
“Medical” (distributed application use case)
Distributed databases, owned resourcesSpecial security needs: privacyEase of installation, maintenance and
access
Small, single-purpose, single-VO dedicated grids
An example: the GPCALMA project
CC Workshop May, 27th, 2004
Piergiorgio Cerello ([email protected]) 7
GPCALMA Screening
CAD selection to minimize data transfers
1 - Data Collection
2 - Data Registration
3 - Run CAD remotely
4 - Transfer Selected Data
5 - Interactive Diagnosis
Data Catalogue
Data Collection Centre Diagnostic CentreData & MetaData Catalogue
Data Catalogue
CC Workshop May, 27th, 2004
Piergiorgio Cerello ([email protected]) 8
GPCALMA Tele-training & Epidemiology
1 - Data Selection
4 - Remote Analysis
3 - Spawn Processes
5 - Retrieve & Analyze Selected Images
2 - Start CAD
Data Catalogue
CC Workshop May, 27th, 2004
Piergiorgio Cerello ([email protected]) 9
GPCALMA CAD on demand
3 - Ask for CAD 4a - Transfer Image4b - spawn PROOF process
1 - Data Acquisition2 - Data Registration
ComputingElement
StorageElement
Data Catalogue
ComputingElement
5 - Run CAD algorithm 6 - Send CAD results
CC Workshop May, 27th, 2004
Piergiorgio Cerello ([email protected]) 10
How to implement the above described Use Cases?
Move code rather than data Share the images without moving them
Single VO in hospitals Secure Access Distributed Data Management Scheduling of Computing Resources
GPCALMA
PROOF ( http:// root.cern.ch )
AliEn ( http:// alien.cern.ch )
CC Workshop May, 27th, 2004
Piergiorgio Cerello ([email protected]) 11
The GPCALMA Graphic User Interface
In use:BariNapoliPisaSassariTorino
CC Workshop May, 27th, 2004
Piergiorgio Cerello ([email protected]) 12
The GPCALMA distributed system configuration
gpcalma.to.infn.it
Server
Distributed System Configuration Users’ Database Data Catalogue Web Portal
Node
Client
Node
Client
Node
Client
Client Storage Element File Transfer Daemon ROOTd/PROOFd GPCALMA
Node
Client
Node
Client
Clients installed: Lecce, Napoli, Pisa, Sassari, Torino
CC Workshop May, 27th, 2004
Piergiorgio Cerello ([email protected]) 13
The AliEn-GPCALMA Core Serviceshttp://gpcalma.to.infn.it
CC Workshop May, 27th, 2004
Piergiorgio Cerello ([email protected]) 14
Patient creation
Image registration
Catalogue query
CC Workshop May, 27th, 2004
Piergiorgio Cerello ([email protected]) 15
The basic functionality is available and tested
Demos presented at SC2003 and HG2004
GPCALMA Achievements
Ongoing tasks: C++ ROOT-AliEn API for Input Data Selection improve the algorithms performance – new approaches optimise the implementation of data and metadata set up a prototype in the participating hospitals
CALMA algorithms rewritten in C++, based on ROOT New GUI, with functionality to manipulate the images AliEn server and clients operational PROOF cluster configured 1st mammogram remotely analysed in March 2003 data/metadata structure being (re)defined re-organisation of the CALMA Database CALMA-DICOM format conversion
2002
2003
CC Workshop May, 27th, 2004
Piergiorgio Cerello ([email protected]) 16
GPCALMA CAD News Masses
ROI Search
Features AREA Perimeter/AREA Entropy Fractal Dimension
Neural Network
ENTROPIA
00,0050,01
0,0150,02
0,0250,03
0,0350,04
0 20 40 60 80 100
SANI
MALATI
RAPPORTO PERIMETRO AREA
0
0,01
0,02
0,03
0,04
0,05
0,06
0,07
0,08
0 20 40 60 80 100
SANI
MALATI
AREA
0
0,05
0,1
0,15
0,2
0,25
0,3
0 20 40 60 80 100
CURVA ROC
0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
1
0 0,2 0,4 0,6 0,8 1
CAD contorni CAD cerchi
CC Workshop May, 27th, 2004
Piergiorgio Cerello ([email protected]) 17
GPCALMA CAD News Microcalcifications
1.H
-Dom
e R
eco
nstru
cted
Imag
e2
.Maske
d
Imag
e3
. Ob
tain
ed
Bin
ary
Im
ag
e4. C
onnecte
d C
om
ponen
ts Labellin
g
Image
CC Workshop May, 27th, 2004
Piergiorgio Cerello ([email protected]) 18
GPCALMA CAD News Microcalcifications
Pre-Processing
Features AREA Perimeter/AREA
Neural NetworkClassification: negative,
benign, malignant
Number of samplesNN not reached False Clusters Benign Malignant
False Clusters 29 6 19 4 0
Benign 5 0 1 4 0
Malignant 8 0 0 1 7
CC Workshop May, 27th, 2004
Piergiorgio Cerello ([email protected]) 21
MAGIC-5 INFN expertise and leadership in:
CAD development Grid Middleware
Does any other Medical field but mammography require a similar approach? CAD for Lung Cancer detection… it’s on time – like CALMA!
3D CT images search for different patterns same Grid approach
AliEn is presently the best available Grid implementation in terms of easiness of installation, functionality, stability and scalability
Alzheimer’s disease diagnosis Colonoscopy (?)
MAGIC-51 project (MAGIC-5) and common GRID Services3 Virtual Organisations
GPCALMA ANPI (Analisi Neoplasie Polmonari in Italia) ADD (Alzheimer’s Disease Diagnosis)
MAGIC-5
ADDGPCALMA ANPI COLON
CC Workshop May, 27th, 2004
Piergiorgio Cerello ([email protected]) 22
CAD for Lung Cancer? 5 years survival rate for lung cancer:
14% (US), 10-15% (EU) no improvement in the past 20 years
Low dose CT: 6 times more efficient than Chext X-Ray (CXR) in the detection of state I malignant nodules
CAD methods are being explored
Gurcan et al., Med. Phys. 29(11), Nov. 2002, 2552: “…computerized detection for lung nodules in helical CT images is promising…large variations in performance, indicating that the computer vision techniques in this area have not been fully developed. Continued effort will be required to bring the performances of these computerized detection systems to a level acceptable for clinical implementation.”
Number
of cases
Sensitivity
(%) FP/image Authors
17 95.7 0.3 Fiebich
17 72 4.6 Armato
26 30 6.3 Fiebich
43 71 1.5 Armato
16 86 2.3 Ko
34 84 1.74 Gurcan
About 43 images/patient About 0.5 MB/image
CC Workshop May, 27th, 2004
Piergiorgio Cerello ([email protected]) 23
Linear patient motion through the gantry Beam rotation
spiral pattern of data acquisitionone continuous set of volume dataReconstruction options
(Slice reconstruction increment) (Interpolation algorithm) (Effective slice thickness)
Spiral CT imaging principles
Best available trade-off between sensitivity for the detection of nodules and
absorbed dose
Single(Multi)-slice: 1(1) tube + 1(N) detector array(s) with 500-900 elements + 1(4) DAQ channel: 1(2)D curved array, shorter scan time N >= 4 detector arrays
(A)symmetric detector arraysDetector elements or arrays can be combined to obtain different thickness and/or widthCollimators can also be used
CC Workshop May, 27th, 2004
Piergiorgio Cerello ([email protected]) 24
Multi-slice vs. Single-slice Volume Coverage:
N x P x S x T
RN= number of DAQ channels = 4P= pitch (linear movement in T/beam collimation)
S= detector width (mm)T= execution time (s)R= rotation time (s) = 0.5 s
mAs kV Collimation (mm) Pitch T (s) Step (mm)
SSCT 43 140 3-5 2:1 1 1
MSCT 20 120 1(x4) 7:1 0.5 2-5
CC Workshop May, 27th, 2004
Piergiorgio Cerello ([email protected]) 25
Images: an example
5 mm 140 KV 120 mAs
Ric. 5 mm 120 KV 20 mAs 1 mm 120 KV 20 mAs
CC Workshop May, 27th, 2004
Piergiorgio Cerello ([email protected]) 26
Screening in Italy & EU-US Main goal
reduce the death rate caused by lung cancer The sample55-69y>20 (packs/day) * ySmokers (or ex-smokers < 10 y)AgreementNo previous cancer
ItalyOngoing programs: Genova, Milano, Torino Starting phase: Regione Toscana – Emilia-Romagna
About 7000 exams in 4 years EU – US
Collaborative Spiral CT-groupI-ELCAP: International Early Lung Cancer Action Project EU ELCDG: EU Early Lung Cancer Detection GroupUS: National Lung Screening Trial (50,000 people)
CC Workshop May, 27th, 2004
Piergiorgio Cerello ([email protected]) 27
Interface for GRID applications Statistical analysis of PET images databases for the study of the Alzheimer Disease
Alzheimer Disease (AD) is the leading cause of dementia, accounting for more than half of all dementias in elderly people
Why Grid?Highly difficult collection of a control group built with normal images
Remote access to a database of normal patients
Access control (Cfr registration, autentication, certification)
Interactive SPM Statistical Analysis
Neuroinformatics Portal
Minimal statistical valueMinimal statistical value 15 SUBJECTS1 SUBJECT 10 MIN
Best statistical valueBest statistical value 150 SUBJECTS1 SUBJECT ?
CC Workshop May, 27th, 2004
Piergiorgio Cerello ([email protected]) 29
SET of CONTROLS 1(PET, SPECT IMAGES)
STATISTICALSTATISTICALTOOL (SPM)TOOL (SPM)
SET of CONTROLS 2(PET, SPECT IMAGES)
SET of CONTROLS 3(PET, SPECT IMAGES)
SET of CONTROLS n(PET, SPECT IMAGES)
PORTAL
UPLOAD
IMAGE of PATHOLOGIC
SUBJECT(PET or SPECT IMAGE)
STATISTICALANALYSIS
OF THE UPLOADED
IMAGE
The Alzheimer Diagnosis Use CaseUniv. Ge, MiB, Osp. S. Raffaele
CC Workshop May, 27th, 2004
Piergiorgio Cerello ([email protected]) 30
SPM ClientData Collection
Portal
AliEn Server
SPM Server
DB Catalogue
PROOF Master
Repository Node
AliEn Client
SPM Server
DB Reference
Data collection
Root Client
Alzheimer Disease Use Case
Server
Repository Node
AliEn Client
SPM Server
DB Reference
Data collection
Root Client
Repository Node
AliEn Client
SPM Server
DB Reference
Data collection
Root Client
SPM ClientData Collection
Server NodeUser Node
User Node
CC Workshop May, 27th, 2004
Piergiorgio Cerello ([email protected]) 31
Alzheimer Disease Use Case
Image Normalisation
Data catalogue Query
Image Transfer
Statistical Analysis
Maps Transfer
Image Normalisation
Image Comparison
Results Transfer
Repository Node
Image Acquisition Reference Atlas Selection
Image Transfer
Maps Visualisation
Server Node
User Node
1
2 2
33
4
Image Normalisation
Image Comparison
Results Transfer
Repository Node
CC Workshop May, 27th, 2004
Piergiorgio Cerello ([email protected]) 33
Conclusions Breast Cancer Detection in Screening Programs: good example of e-health
application that would benefit from the use of GRID Services
The AliEn/PROOF based approach allows:Minimisation of data transfers Secure management of a distributed Virtual Organisation
The success will depend on: the reliability and stability of interactive GRID Servicesthe performance of CAD algorithms: ongoing new approaches the quality of the GUI
GPCALMA Virtual Organisation in the participating Hospitalsby the end of 2004 with improved CAD algorithms
New applications will followANPI, ADD, COLON
EGEE/LCG/ARDA: Architecture Roadmap towards Distributed AnalysisPrototype developed in the framework of EGEE by Sep 2004Migrate to that prototype
CC Workshop May, 27th, 2004
Piergiorgio Cerello ([email protected]) 34
UserInterface
InformationSystem
ComputingElement
Data & MetadataCatalogues
StorageElement
AuthenticationAuthorisation
MonitoringAccounting
“Green” VO
“Blue” VO
WM
S