Grid Analysis of Radiological Data

Grid technologies and infrastructures can contribute to harnessing the full power of computeraided image analysis into clinical research and practice. Given the volume of data, the sensitivity of medical information, and the joint complexity of medical datasets and computations expected in clinical practice, the challenge is to fill the gap between the grid middleware and the requirements of clinical applications.This chapter reports on the goals, achievements and lessons learned from the AGIR (Grid Analysis of Radiological Data) project. AGIR addresses this challenge through a combined approach. On one hand, leveraging the grid middleware through core grid medical services (data management, responsiveness, compression, workflows) targets the requirements of medical data processing applications. On the other hand, grid-enabling a panel of applications ranging from algorithmic research to clinical use cases both exploits and drives the development of the services.
 INTRODUCTION Harnessing the full power of computer-aided image analysis into clinical research and practice remains an open issue. Given the amount of data produced by X-ray Computed Tomography (CT), Magnetic Resonance Imaging (MRI), or PET-scan, and the difficulty to interpret medical images, algorithms for medical image analysis, processing, and diagnostic assistance have been developed these last 15 years or so. Some of these algorithms have reached a high level of usability and proved to have a real impact in the clinical domain. However, their widespread adoption by clinicians is not realized yet. Two stringent examples, amongst many others, are radiotherapy, which could greatly benefit of exploiting advances in segmentation and registration algorithms, and intra-operative situations as well as intervention planning, which could exploit modern high-performance computing systems for augmented reality (Kikinis, 1998). In the 90’s, G.A Moore has coined the term “crossing the chasm” (Moore, 1991) for the issue of selling hightech products to mainstream customers. Crossing the long-lasting chasm between, on one hand the advances in computer science and engineering in the field of medical images analysis, and on the other hand clinical research and practice, is a challenge of the same nature. This chapter reports on the goals, achievements and lessons learned from the AGIR (Grid Analysis of Radiological Data) project. AGIR is a multi-disciplinary collaboration funded by the French ministry of research under the ACI/ANR scheme, for the 2004-2007 period. AGIR gathers collaborators from eight laboratories, including computer scientists, middleware experts and physicians. The central tenet of AGIR is that grids can help crossing the chasm, despite their recent apparition, because they introduce a change in paradigm in the access to high-end resources. Grids provide seamless scaling of the development, validation, and exploitation cycles of analysis methods: the same infrastructure allows multiple user communities to access and manipulate medical data, and to explore individual images or create augmented reality situations through compute-intensive visualization methods; the same infrastructure provides the computing power needed to validate algorithms on large datasets and to process complete databases for the most demanding applications such as epidemiology. Filling the gap between the clinical applications and the grid middleware raises many specific issues, ranging from computer science basic research to legal concerns, as will be exemplified in section 2. Addressing all these issues is an active research and technology area, and a new scientific community is emerging (Berry, 2003). In this framework, the specific objectives of AGIR are to define and validate: • new grid services that address some of the requirements of complex medical image processing and data manipulation applications; these services are described individually in section 3. • new medical image processing algorithms taking advantage of the underlying grid infrastructure for compute and data intensive needs; sections 4 to 6 report on these developments. The method is to confront the expertise of medical and computer science teams, specialized in clinical applications, medical images analysis algorithms, grids and distributed systems, around a few paradigmatic medical applications, in order to get a cross-section of the middleware, algorithmic and medical issues. A specific application (section 7) was a “bottom-up” approach targeting the immediate development of a telemedicine platform, whose requirements and results were to be confronted with the more long-term activities in AGIR. GRID-ENABLING MEDICAL IMAGE ANALYSIS Challenges and issues Grid computing has been considered to tackle some of the requirements of the medical image analysis community (Montagnat, 2004; Aloiso, 2005). The scenarios investigated in AGIR (see section 4-6) led to the identification of common area-specific grid services: • Medical databases management in order to handle large amounts of sensitive data. Federation of distributed data sources is needed e.g. for large scale testing involved in algorithms validation. • Optimized and reliable data transfers even on low-end networks, e.g. to support humanitarian medicine. • Efficient execution of analysis pipelines on a distributed grid infrastructure in order to match the computing requirements of the most demanding applications such as cardiac sequences analysis. • Responsiveness needed by human guided procedures such as assisted segmentation and surgery planning. • Production quality support to ensure sustainable grid usability. Overview of AGIR The deployment of medical applications on grids can be structured following four layers: 1) general purpose grid services; 2) the core services dedicated to medical applications that are not available in general purpose middleware; 3) the medical image processing algorithms taking advantage of the underlying grid services to process large amounts of data or to run expensive computations; 4) clinical applications developed to address medical challenges. Figure 1: AGIR structure AGIR mainly addresses core medical services and image processing algorithms and, to some extent, clinical applications. Figure 1 summarizes the various AGIR activities. einfrastructures Grid concepts and research are around since the 90’s at least (Foster, 2003) (Foster, 2001). With the maturation of technologies, and the deployment of production grids, however, the focus has now shifted towards strategic issues, amongst which forecasting and shaping roadmaps towards future e-infrastructures (Atkins, 2003). Following the analysis of the e-IRG (e-infrastructures reflection group, a body of EU member states representatives advising the European Commission), sustaining the needs of science requires resource integration at the European scale not as a fully integrated (“monolithic”) grid, but at the level of an interoperable grid of grids (Leenars, 2005). At the beginning of AGIR, these issues were not fully formulated. The popularity of the Globus middleware (Foster, 2006) pushed towards autonomous developments. However, we already considered that it was critical for AGIR to refrain from developing middleware functionalities, which were already available, even if not perfectly suited for its needs. Throughout the project, we tried to keep a balance between neutrality with respect to middleware implementations on one hand, and strategic concerns on the other hand. On the generic side, we conducted experiments and analysis in order to assess the compatibility of various middleware platforms, from research ones to production oriented ones. Amongst the numerous research suites, our experiments focused on DIET (The Distributed Interactive Engineering Toolbox) (Caron, 2006), which exemplifies remote procedure call (RPC)based architectures. Another important component of our experiments is the French national grid, Grid’5000 (Bolze, 2006), which allows repeatable experiments in controlled settings. On the strategic concerns side, AGIR has developed a close collaboration with the FP6 and FP7 infrastructure projects EGEE (Gagliardi, 2005), which is expected to move towards a permanent European Grid Infrastructure federating National Grid Infrastructures (NGI). EGEE is strongly involved in the various prospective bodies such as e-IRG, interoperability actions with other international grids (e.g. NorduGrid or Open Science Grid), and with the standardization Middleware evaluation Medical data Management Medical data access protocols Responsiveness Workflow Management Cardiological Segmentation Registration in oncology Volume reconstruction Humanitarian Medicine efforts of the Open Grid Forum. EGEE is thus one of the key actors in the European einfrastructures theater. EGEE is a controlled but widely distributed environment over the Internet which usage for manipulating sensitive medical data imply the implementation of strong protection measures. However, the strong interaction with EGEE has a triple advantage. The first and most immediate one is the sheer availability of resources, showing that IT resources are readily available, which is a cornerstone for real-world usage scenarios. The second advantage is to interact with middleware developments, specifically gLite (Laure, 2004). On one hand, this process is considerably more difficult than it would be in an experimental grid. On the other hand, the impact of these requirements, when satisfied, benefits from the general impact of EGEE, and are sustainable. An alternative might be to separate grids for health from other e-infrastructures. In the AGIR time frame, other large scale infrastructures targeting life sciences have been deployed worldwide, adopting this strategy. In particular, the BIRN (Biomedical Informatics Research Network) project (Peltier, 2003) has deployed an homogeneous hardware and software infrastructure all over the USA. Similarly on the EU side, Health-e-child 1 has deployed an EGEE-compatible but independent infrastructure. This has the advantage to minimize the legal and ethical issues as the nodes are owned and controlled by the hospitals themselves so that they comply to each site requirements. Moreover, the different sites are much more easily administrated by the consortium. We believe that this is a good way to expose the clinical community to grid technologies in order to facilitate in the future the acceptance and the use of more standard einfrastructures. Conversely, the CaBIG 2 project has adopted a more open approach with the publication of different level compatibility recommendations for the different components, existing or to appear, of its grid infrastructure. In both cases the logic of a production grid is the same: achieving a proper balance between, one hand quality and sustainability (24x7 availability, and compliance to standards), and on the other hand coping with scaling requirements, both in heterogeneity (workloads and use cases), and in geographical and institutional distribution. CORE MEDICAL SERVICES Grid interface for Medical Data Management Medical data management in clinics is a complex problem given the multiplicity of data sources, the sensitivity of the data, and the tremendously large data sets involved. As a consequence, medical information systems are often fragmented between different clinical sites where homogenization efforts are rarely invested. Even inside a single health enterprise, it is not rare to encounter several information systems partially overlapping and covering the different aspects of clinical information with more or less success. Considering medical images, we have developed a complete solution, the Medical Data Manager (MDM). The MDM, interfaced to the EGEE middleware, bridges the DICOM storage used internally to clinical centres and the grid storage units (Montagnat, 2008). The MDM both ensures a transparent access to medical images by any grid service and does not interfere with the usual clinical workflow of images. Benefiting from the distributed storage capability provided by the grid middleware, it enables data sources federation: large image sets originating from different sites can be assembled for studies involving processing population of images such as the algorithm validation through statistical procedures as the Bronze Standard (section 5). 1 Health-e-child: Integrated platform for paediatrics, http://www.health-e-child.org/ 2 CaBIG: Cancer Bioinformatics Grid, https://cabig.nci.nih.gov/