Level 2: Grid-enabled image-analysis, image normalisation and acquisition error trapping
2.1: The NeuroGrid image analysis toolkit (an algorithm wrapping portal)
Members of this consortium have developed a wide range of image analysis algorithms, many of which are automatic and have been applied in multiple neuroimaging research projects. We will combine these into a neuroimage analysis toolkit that can be used throughout the consortium, and in due course disseminated to other researchers. Because these algorithms are typically mutually incompatible in terms of the required hardware, operating system and input/output formats, the construction of a unified toolkit in the traditional way would involve substantial re-implementation and validation, and would require enormous investment - along the lines of the multimillion dollar programme of US NIH National Library of Medicine for the Insight toolkit. In practice, it would be much more efficient to work in a distributed heterogeneous environment. e-Science offers an efficient and flexible means of building a toolkit based on existing C and C++ algorithms presented as Grid services. This concept has already been demonstrated by IXI using individually customised solutions that can then be accessed via a unified interface. Each service can be run remotely on different hardware as required including the use of condor clusters to provide high throughput. Additional services can be implemented that convert between input and output file formats and the services can be combined using simple workflows described using XML.
We will implement an algorithm-wrapping portal within the NeuroGrid project that will greatly simplify this process of making an existing algorithm available as part of a computational Grid. To add an additional algorithm to the Grid-toolkit, a researcher would go to the portal and follow the instructions to upload their source code to a suitable remote machine or install globus on their local machine, specifying all the parameters required by the algorithms. The necessary java code and XML service specifications needed to define the service would then be automatically generated. The algorithm could then be registered with an index service, thus becoming a discoverable application that can be used both by the researcher who submitted it, and by the rest of the consortium.
The initial NeuroGrid image analysis toolkit will comprise the image analysis services generated by IXI. We will test the portal capabilities by wrapping the Institute of Neurology Dementia Research Group algorithms and make these available across the consortium. When the system is proved, it will be documented and training courses will be run to assist others use the facility. The core toolkit development team will work to empower other consortium members to create new services using the portal as required for the exemplars. The NeuroGrid-image analysis toolkit thus generated will be combined with a simple workbench that enables users to discover the services available within the project (both image analysis algorithms and file format conversion routines), and provide templates to combine these in simple workflows to meet their application requirements. We will also link the services that have been created in this way to the databases used in the NeuroGrid project and investigate how services created in this way can be incorporated into data provenance and audit trails.
The toolkit will offer some very attractive features because the use of these image analysis services will become as easy as using a web page, unlike the current task of down-loading software from the web and getting it running on a suitable local machine. It will become straightforward to compare different algorithms with the same functionality (e.g. if there are multiple algorithms in the toolkit that register images, segment MR scans into tissue classes, or quantify cerebral atrophy, it will be simple to apply different algorithms to a given set of data and compare the results) and it will become easy to build an application that makes use of a collection of algorithms that could not normally be used together because they run on different systems.
2.2: Image Normalization
When different scanners are used, the images produced can have significant differences in contrast and point-spread function that are associated with difference in manufacture and physical specifications, and which cannot be eliminated by quality control measures. A major focus in the past has been size scaling and this can get mixed up with spatial normalisation between subjects. There has been a long history of attempts to deal with this using QA methods, mostly based on phantom objects or human phantoms who visit each scanner. There has also been use of image registration to normalise to atlas or to reference subjects. There are shared issues across modalities, but also many modality specific- e.g. Gradient scaling in MRI, Gantry calibration and tilt errors in CT etc.
Another issue is that different scanners from different manufacturers can give different contrast from nominally the same scan, and this is a major factor in MRI. There have been few coherent efforts to deal with this problem which remains largely unsolved. One approach is quantitative MRI with determination of basic parameters, pd, T1, T2 so that any contrast can be generated, but this has proved a poor way to mimic the native contrast of T1 and T2 weighted scans and results in long examination times for each patient.
An advantage of an e-Science approach to imaging studies is that, by aggregation of large amounts of image data, we can learn scanner variability from the data itself. By non-rigidly registering all the scans together, we can warp all the scans to make them look virtually identical. The residual variability will be due to (a) errors in the registration and (b) variation in intensity due to different scan contrast. Errors in registration tend to be concentrated around the cortex, where inter-subject anatomical variability is greatest. The variability in the white matter and basal grey matter is, therefore, predominantly due to the variation in the scanners, and we can fit a model to this variability which we can use to compensate for the variability for the entire images. This approach can be used as a "calibration" step on human phantoms carried out before the start of a study, or can be used once all the data has been collected to additionally compensate for variation in scan properties that may result from scanner upgrades during the course of the projects.
NeuroGrid will investigate the potential of using this technique to compensate for contrast variation in imaging studies that have the same scan protocol at all sites (where scan differences will be quite small), and also for the comparison of data from different retrospective studies that might have been acquired with rather different scan protocols (where scan differences may be quite large).
2.3: Catching errors before the patient leaves the examination.
At present patients in multi-centre studies are scanned to fixed protocols with no verification at the time of examination. Simple errors can destroy the value of the data acquired. The web offers the potential to verify data entered into web-forms and this is already having an impact on clinical trials. The Grid offers the potential to verify imaging data using analysis tools to check for data consistency both with the desired protocol (correct regional examination with correct contrast parameters) and also to signal quality assurance errors (sudden decline in signal homogeneity or SNR). This image quality assurance involves both checking the scan metadata (in the DICOM header) to ensure the correct scan parameters (e.g.: TE, TR, field of view, resolution), and also comparing the image voxel data itself against previous scans of the same subject and against other subjects in the cohort to ensure the image intensities have the correct characteristics, that the images are not corrupted by artefact, and that the correct part of the patient has been scanned. This second task is extremely computationally demanding as the algorithms concerned are CPU intensive, the scans of about 10Mbytes will need to be transferred across the network, and comparisons will be made with reference data (potentially multiple images each of a similar size) that may be distributed across multiple sites. We propose to explore the options to use suitable Grid services to validate data on the fly during examinations, so that errors can be corrected and variance between sites and between examinations can be reduced. Our aim will be to check that the scans are suitable within about 5 minutes. Some of the scanners used in this project are university owned research scanners that are directly connected to a university network, and can "export" data to the Grid services over a the network. However, as in most multicentre studies, many scanners available to this consortium are run by NHS hospitals and are on separate physical networks to the Grid services. For these centres we will install a workstation in the console room, connected to the university network, which can read the removable media generated by the scanner. After the first scan has been acquired, it will be written to removable disk (probably CD), walked across the room to the Grid workstation, and uploaded to the Grid for the quality assurance.
