CC Workshop May, 27 th, 2004Piergiorgio Cerello ([email protected])1 MAGIC-5 Medical Applications...

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



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



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



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



UserInterface

InformationSystem

ComputingElement

Data & MetadataCatalogues

StorageElement

AuthenticationAuthorisation

MonitoringAccounting

“Green” VO

“Blue” VO

WM

S



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



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



GPCALMA Tele-training & Epidemiology

1 - Data Selection

4 - Remote Analysis

3 - Spawn Processes

5 - Retrieve & Analyze Selected Images

2 - Start CAD

Data Catalogue



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



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 )



The GPCALMA Graphic User Interface

In use:BariNapoliPisaSassariTorino



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



The AliEn-GPCALMA Core Serviceshttp://gpcalma.to.infn.it



Patient creation

Image registration

Catalogue query



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



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



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



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



GPCALMA from GENIUS



GPCALMA on iBook



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



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



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



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



Images: an example

5 mm 140 KV 120 mAs

Ric. 5 mm 120 KV 20 mAs 1 mm 120 KV 20 mAs



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)



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 ?



SET of CONTROLS 1(PET, SPECT IMAGES)

STATISTICALSTATISTICALTOOL (SPM)TOOL (SPM)



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



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



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



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



UserInterface

InformationSystem

ComputingElement

Data & MetadataCatalogues

StorageElement

AuthenticationAuthorisation

MonitoringAccounting

“Green” VO

“Blue” VO

WM

S

CC Workshop May, 27 th, 2004Piergiorgio Cerello ([email protected])1 MAGIC-5 Medical Applications...

Documents

Transcript of CC Workshop May, 27 th, 2004Piergiorgio Cerello ([email protected])1 MAGIC-5 Medical Applications...