Project Background
The computational requirements for Medical Image Analysis are increasing rapidly, with the desire to acquire and extract more complex information from image data sets. These data may consist of many sets of images, for example, images may be acquired rapidly over a period of time to study how an administered contrast agent behaves as it passes through a particular organ. While a single multi-slice data set may be around 10 Mb in size, more complex temporally acquired data may reach several hundred megabytes. Increasingly, an imaging exam of an individual subject will consist of several of these complex data sets each looking at a different aspect of anatomy or physiology, resulting in a data size that may run to over a gigabyte per patient. For a research imaging centre that performs complex imaging studies consisting of many patients, perhaps imaged serially to track disease progression, the storage requirement quickly runs into terabytes.
Figure 1. Structural MR image examples illustrating T1-, T2- and PD-weighted image contrast.
Furthermore, the more sophisticated imaging data require substantial computational processing, as data sets need to be combined and manipulated in order that quantitative imaging parameters can be extracted from the raw imaging data. This typically involves motion correction and temporal registration of the data, and then mathematical models are fitted to the data on a voxel-by-voxel basis to produce quantitative parameters that represent different aspects of anatomy or physiology. The latter processes produce significant computational load, as a typical data set may consist of anywhere between 250,000 and 50 million voxels. There is significant scope for developing efficient parallel implementations of these processes that would reduce the time required for data processing and enable the information obtained from these techniques to be relayed back to the clinician in 'real time' rather than being processed off-line and presented the following day.
Figure 2. Quantitative images derived from raw MR image acquisitions using computational post-processing techniques.
Once images have been processed, relevant clinical information must be extracted. Traditionally, this takes the form of an output image being presented via film or monitor and interpreted visually by the clinician. Increasingly, however, quantitative data is required to be extracted from the output images. This may take the form of the location and volume of regions of suspicion, or it may be that values representing a particular aspect of physiology are to be extracted from a region-of-interest. Increasingly, the desire is for these processes to be automated, or at least semi-automated, to reduce the burden on the user and facilitate 'computer-aided diagnosis'. In more sophisticated studies, this data may need to be extracted into a database, so that it can be subjected to further statistical analysis. The post-processing procedures applied to the data need to be stored along with the considerable meta-data generated en-route, so that it may be interpreted and the results reproduced at some time in the future.
It is clear that modern imaging data needs to be efficiently managed, accessed, and processed; all of which require Research Computing methodologies to be developed, implemented and optimized.
Further background information can be found at the SFC Brain Imaging Research Centre website: http://www.sbirc.ed.ac.uk/.
Publications
See Publications