MARS - Models of Appearance, Registration and Segmentation

E may soon start working in collaboration with the MIAS-Grid, as well as the IRC. MIAS-Grid, where MIAS stands for Medical Images and Signals, is a project which ultimately produces an e-Science workbench for medical image analysis. To demonstrate the utility of the system, a series of use-cases is required and our code might be among these.

Essentially, the Grid might have our registration assessment algorithms re-created. It will then compartmentale the processes and carry out some analysis in a transparent way that has robust, well-understood interfaces (e.g. XML-RPC). Subsequently, these processes can be embedded as workflows within the workbench, which might involve autonomous and powerful computers. Algorithms can easily be exchanged, thus enabling benchmarks (comparisons) to be rapidly conducted. The idea is reminiscent of the principles and rationale behind the Strategy Patten in OO programming.

Image registration assessment: the benchmark architecture
Click image for full-sized version

We are not too certain about the future of this initiative and, in particular, some of the technicalities. Yet, we would feel privileged to have an opportunity to work on such modern computing architectures. Our particular set of binaries can directly benefit from parallel workflows. In short, here is the framework that can be envisioned already:

We are given a set of N images
We have M such image sets
We need to build a model for each set among these M sets. That can definitely be done in parallel and there is no apparent dependency
We now proceed to evaluating M models and again there are no dependencies among the evaluation processes

To add some context, registration is evaluated through the construction of appearance models. All in all, the process in question need not be serial and it can be handled merely (not entirely) in parallel. We can further refine speeds by treating sub-sets of data (chunks) and then aggregating the results, if needed.

This would be similar to things we have done in the past, such as deploying banaries in computer clusters, invoking them via SSH, and collecting the output later. At extremity we used 30 units overnight to produce some urgently needed results.