The MammoGrid Project

The MammoGrid ProjectThe MammoGrid Project- an EU FP5 funded project

On behalf of the MammoGrid Consortium:CERN, Geneva; Mirada Solutions, Oxford;

Universities of Oxford, Pisa, Sassari, West of England;

University Hospitals, Cambridge (Addenbrooke’s)& Udine (Policlinico Universitario)

Tony Solomonides, UWE

Tony [email protected] [email protected] XIX NEC 15-20 September 2003XIX NEC 15-20 September 2003 page 2 / N

Key StaffKey Staff

University of Cambridge (Addenbrooke’s Hospital)Ruth Warren (Radiology)

CERNRoberto Amendolia (Project Leader), Predrag Buncic (AliEn), Josè Galvez

Mirada SolutionsRalph Highnam (SMFTM), David Schottlander (MAS)

University of OxfordMike Brady, Chris Tromans (Quality Control)

University of PisaPasquale Delogu, Evelina Fantacci (CADe)

University of SassariUbaldo Botigli, Piernicola Oliva (CADe)

University of West of England, BristolFlorida Estrella, Tamas Hauer, David Manset, Dmitri Rogulin (DICOM, Grid)Richard McClatchey (Technical Director)Mohammed Odeh (User Requirements)Tony Solomonides (User Requirements, Database, Dissemination)

Policlinico Universitario di Universita degli Studi di UdineMassimo Bazzocchi, Chiara Del Frate (Radiology)


ContentsContents

1. Background & Motivationa. Breast screening programmesb. Mammography and its problems

2. Technologies & Gridsa. Mammographic toolsb. CRISTALc. AliEn

3. MammoGrida. Objectivesb. Methods

4. Conclusions & questions


Breast Cancer ScreeningBreast Cancer Screening

• Breast cancer is a major problem:– 12% lifetime risk of breast cancer– 19% of cancer deaths are due to breast cancer– 24% of all cancer cases are breast cancers– in EU & USA 350,000 are diagnosed and

115,000 die annually– but 73% of diagnosed cases survive 5 years or

more

• Early diagnosis improves prognosis


Breast Cancer Screening - EUBreast Cancer Screening - EU

• UK: covers women 50 – 64; established in 1987; 1.3m women screened annually at 92 centres (230 radiologists)

• Sweden, Finland, The Netherlands, Ireland, France and Germany have or are about to establish national programmes

• Italy: depends on where you live; expertise and areas covered by screening programmes do not necessarily correlate

• Variable availability and coverage in other countries


Breast Cancer Screening - UKBreast Cancer Screening - UK

• Low tech, ‘rapid, high-throughput’ screening• Basic frequency every three years• Ideally, a full ‘series’ taken at screening consists of

two images (cranio-caudal and medio-lateral oblique) for each breast – four in all

• Historically, only two CC images taken• A radiologist may have 40 seconds to study the

series• A study in 2000 estimates a 6.4% contribution to

reduced mortality to the screening programme; better treatments account for 14.9%

• More ‘interval cancers’ than anticipated• Pressure is mounting to begin screening every 2

years


Breast Cancer Screening - ProblemsBreast Cancer Screening - Problems

• Sensitivity and specificity – i.e. percentages of false negatives and false positives – are both unacceptable

• In the US, 80% of biopsies are benign – error on the side of caution

• A combination of computer-aided detection and radiologist screening improve performance significantly

• There is a world-wide shortage of trained radiologists and radiographers (radiologic technicians)

• The work is perceived as ‘boring but risky’; in US 12% of malpractice lawsuits are against radiologists

• Over 20% of films are practically unavailable for re-use or comparison – misfiled, in transit or lost


Mammography - ProblemsMammography - Problems

• Image quality (brightness/visibility and contrast/distinction) are affected by tube kVp and mAs, screen and film characteristics and automatic exposure control

• The interpretation of images is difficult, there are issues both about training and about experience

• There are ‘good practice’ guidelines, but radiographers have some latitude in the decisions they make

• In both UK and Italy quality control issues have arisen– Failure to record settings– Inconsistency of settings– Unusable images

• MammoGrid is concerned with film-screen systems; the move to all-digital mammography is slow


What is MammoGrid?What is MammoGrid?

• EU project to prefigure a pan-European distributed database of mammographic images using GRID Technologies.

• Aim: To provide a demonstrator for use in epidemiological studies, quality control and validation of computer aided detection algorithms.


MammoGrid ObjectivesMammoGrid Objectives

1.1. To To evaluate current Grids technologiesevaluate current Grids technologies and determine the requirements for Grid-compliance in a pan-European mammography database.

2.2. To To implement the MammoGrid databaseimplement the MammoGrid database, , using novel Grid-compliant and using novel Grid-compliant and Federated-Database technologiesFederated-Database technologies that will provide improved access to distributed data and will allow rapid deployment of software packages to operate on locally stored information.

3. To deploy enhanced versions of a standardization systemstandardization system that enables that enables comparison of mammogramscomparison of mammograms in terms of intrinsic tissue properties independently of scanner settings, and to explore its place in the context of medical image formats (DICOM).

4. To develop software tools to automatically extract image informationsoftware tools to automatically extract image information that can be used to perform quality controls on the acquisition process of participating centers (e.g. average brightness, contrast).

5. To develop software tools to automatically extract tissuesoftware tools to automatically extract tissue informationinformation that can be used to perform clinical studies (e.g. breast density, presence, number and location of micro-calcifications) in order to increase the performance of breast cancer screening programs.

6. To use the annotated information and the images in the database toto benchmarkbenchmark the performance of the softwarethe performance of the software described in points 3, 4 and 5.

7. To exploit the MammoGrid database and the algorithms to propose initial to propose initial pan-European quality controlspan-European quality controls on mammographic acquisition and ultimately to provide a benchmarking system to third party algorithms.to provide a benchmarking system to third party algorithms.


MammoGrid PhilosophyMammoGrid Philosophy

• Project concentrates on applying emerging GRID technology rather on developing it.

• It plans to implement a ‘lightweight’ (but fully functional) GRID and study its usage in hospitals

• It will draw heavily on other Grids projects e.g. DataGrid

• It will deliver a prototype federated database of mammograms in hospitals in the UK and Italy

• It will provide rapid feedback from the Hospital community

• And will inform the next generation of health grids developments


Why a Mammography Why a Mammography Database?Database?

Improved reliability of screening and early diagnosis requires:– better epidemiological understanding– improved diagnostic tools– enhanced quality control– continuous training– efficient management of data and records.

• Need to establish research and training repositories that contain sufficiently large statistical samples:– MammoGrid-EU– NDMA-US– eDIAMonD-UK– GPCalma-Italy


TechnologiesTechnologies

1.Mammography– SMF™ (Mirada)

– CADe (CALMA)

– DICOM (Medical Imaging Standard)

2.Distributed computation– CRISTAL Database (CERN/UWE)

– AliEn GRID (CERN)


Standard Mammogram FormStandard Mammogram Form


The theory of SMF™The theory of SMF™

• Mirada’s Standard Mammogram Form (SMF™) measures the column of non-fatty tissue between the compression plate and the imaging surface.

• SMF algorithm models the physics of image formation, including extrafocal radiation, scatter, grid effects, film-screen characteristics, etc.

• The contribution of the imaging system is factored out.

• The image is decomposed into fatty tissue and non-fatty tissue.

• The new representation gives a numerical value for the amount of non-fatty tissue at any point on the image.


A normalized imageA normalized image


CMS DetectorCMS Detector

ECAL


ECAL BarrelECAL Barrel


ECAL productionECAL production

SIC (China)Crystals

Bogorodisk(Russia)Crystals

IPN Lyon(France)

Electronics

LPNHE(France)Alveoli

ENEA/INFN (Italy)

Modules

CERN (Switzerland)

Modules & Supermodules

CERN (Switzerland)

SM in testbeam

CERN (Switzerland)ECAL in CMS

Parts

CEA/DAPNIA (France)

Monitoring

Parts

• Parts are shipped between centres• Must control overall construction


Integration via Meta-ObjectsIntegration via Meta-Objects

PartDefin ition

Com positePartDef E lem entaryPartDef

PartCom positeM em ber

0..n

ActivityDefin ition

Elem entaryActDef Com positeActDef

ActCom positeM em ber

0..n

0..n

0..1

Conditions

0..n 0..n

10..n

10..n

CRISTAL 1 concept for integrated product (structure) and process (workflow) description


CRISTAL Object LayersCRISTAL Object Layers

Item C lass Item D escrip tionC lass

Item Item D escrip tion

Is described by

Is described by

Data Meta-Data

Is an instance of Is an instance of

M odel Layer

Instance Layer

Type Object Pattern

Type Object Pattern

CRISTAL 2 Kernel Architecture


CRISTALCRISTAL

• Data and metadata are versioned asynchronously

– Flexibility to cope with and propagate end-users’ changes

– Clear separation between object lifecycles of the domain:

• Metadata: What has to be done

• Data: What is being done

• Data and metadata models are largely domain independent

• ‘Design patterns’ have emerged from the process of abstraction

• Large volume of data can be managed

• Large distributed data processes can be tracked


DICOM Information ModelDICOM Information Model


MammoGrid Object ModelMammoGrid Object Model


Assessment Object ModelAssessment Object Model


The MammoGrid ChallengeThe MammoGrid Challenge

• Building this repository is not trivial because:– Large numbers of exemplars are required.– Cases must be obtained from many

geographically remote locations.– Data itself is large: 2 breasts × 2 views × 4K

× 4K pix × 2 bytes = 128Mbyte per patient per visit, 1.5M women per year UK, ~ 200 Terabytes in UK alone,

– Acquisition is highly variable, same image may look different depending on machine and parameters. How do you compare?

– Patient privacy and data security are key.– Many relevant items of metadata.


GRIDGRID

Seeking a solution for resource sharing & coordinated problem solving in dynamic, multi-institutional virtual organizations.

Data GRID goal:

“Enable a geographically distributed community of thousands to pool their

resources in order to perform sophisticated, computationally

intensive analyses on petabytes of data”


Layered Grid TechnologiesLayered Grid Technologies

Knowledge Grid

DataControl

DataAbstraction

Data Grid

Information Grid

Distributed databases, streaming, near-line storage, large objects, access mechanisms, data staging …

Metadata, middleware, intelligent retrieval, information modelling, warehousing, workflow …

Data mining, visualisation, simulation, problem solving methods/environments …


AliEn ChallengeAliEn Challenge

AliEn = Alice Environment

“Can we provide, building on top of available public domain and open source components and standards, a functional distributed computing infrastructure to the community of our users which will remain operational even if underlying technologies keep changing?”

Geneva, May 2001

Instead of using Globus toolkit or waiting for DataGRID to deliver re-packaged version of Globus, we decided to try different path and use Web Services and related standards as a backbone of our GRID implementation.


AliEn Services in MammoGridAliEn Services in MammoGrid

• Authentication– User’s credentials checked

• Resource Broker– Job/algorithm scheduling

• Storage Element– Data and file management

• File transfer– Scheduled file transfers


Service Oriented ArchitectureService Oriented Architecture


High Energy Physics vs. High Energy Physics vs. MammoGridMammoGrid

• MammoGrid relies heavily on technologies developed primarily in the field of high energy physics.

• Similarities– Large number of big files – Files can be sensibly organized in directory tree– Need to replicate and move file copies between

sites– Need to execute commands on the node which

hosts data locally

• Difficulties– Complexity of co-working in medical environment– Lack of trained IT personnel – Confidentiality


A GRID Infrastructure is A GRID Infrastructure is idealideal

• To test image-based clinical hypotheses, databasesdatabases must be: Populated by large number of cases Contain large files (1 mammogram ~ 32Mb) Geographically distributed repositories Heterogeneous database formats Need to be accessible to co-workers

• Development / validationDevelopment / validation of medical image analysis solutions demands: Computationally expensive simulations Repeated runs for optimal parameter tuning Statistical test rigs Remote execution and maintenance

• Services (e.g. security) must be system-resident, invisible, Services (e.g. security) must be system-resident, invisible, genericgeneric


Federated System Federated System SolutionSolution

Hospital Italy

Healthcare Institute

University Database

Hospital UK

Shared meta-data

Analysis-specific data•Knowledge is stored alongside data

•Active (meta-)objects manage various versions of data and algorithms

•Small network bandwidth required

Clinician’s Workstations

QueryResult

LocalQuery

LocalAnalysis

LocalAnalysis

LocalAnalysis

LocalAnalysis

Massively distributed dataAND distributed analyses

GRIDLocalQuery

LocalQuery

LocalQuery


AliEn Virtual OrganizationAliEn Virtual Organization


Multi-level Virtual Multi-level Virtual OrganizationOrganization


Overall Grids Overall Grids ArchitectureArchitecture

GridBox

GRID

VPN

WorkstationsWorkstations

MammoGrid Data

MammoGrid Data

MammoGrid Data

Udine

Oxford

CERN

MammoGrid Data

Cambridge

GridBox

GridBox

GridBox

High Security Level

Mirada WST(‘‘MAS’’)

Mirada WST(‘‘MAS’’)


Stack of Grid ServicesStack of Grid Services


Store onLocalStore

Workstation Local GridBox

DICOM File

Extract Meta-Data and

SaveCentrally

ALIEN

Meta-Data

Search Criteria

PREPAREList of Files that W/S is Interested In (LFNs)

Find PFNRETRIEVE

LFN

RemoteRetrieval of

Files to LocalData Store

Cache

Requests toother gridservers

2. QUERYQuery

DatabaseList of matching results, including Logical

File Names (LFN)

LocalLFN /PFN

Mapping

1. AcquireNew

Image

Local Data Store

Local Cache

File

Store

DICOMSCU

DICOMSCP

ADD(Trigger)

DICOMSCU

DICOMSCP

C_STORE

C_GET

PFN (dicom://ip:port:aetitle)

DicomGet

List of LFNs Sorted by Estimated Access Time

MammoGrid TransactionsMammoGrid Transactions


MammoGrid QueriesMammoGrid Queries


Main Deliverables and Main Deliverables and MilestonesMilestones

• User Requirements Specification and Technical System Specification (published)

• Packaged medical imaging workstation with interface to GRID, secure GRID box, (October 2003)

• Grid compliant SMF software (December 2003)

• Prototype GRID-compliant database and information infrastructure (March 2004)

• Application software (under development)

• Clinical Trials (late 2004 – end of project)


Conclusions and Further Conclusions and Further ResearchResearch

• Distributed Health informatics is an important application area for Grids technologies – Health Grids

• Many similarities with High Energy Physics

• MammoGrid user requirements specified, but need rapid feedback from the user community

• Effective Grid deployment needed now

• Many questions subject to further research:

– How to resolve distributed queries ?

– What role for meta-data ?

– How to maintain secure, reliable data ?• MammoGrid: First results expected late 2003

The MammoGrid ProjectThe MammoGrid Project

Thank you


Documents

The MammoGrid Project