NeuroImaging Ontologies

The 11th Intl. Protégé Conference is to be held on June 23-26, 2009 in Amsterdam, Netherlands this year and I wanted to do a quick post on one of the tracks with respect to ontology-driven software development. There is an excellent article by Knublack entitled, "Ontology-Driven Software Development in the Context of the Semantic Web:An Example Scenario with Protege/OWL" at http://www.knublauch.com/publications/MDSW2004.pdf on building applications based on the vision of the Semantic Web that uses OWL and Web Services along with agile development methodologies. As you can see from other posts on this site and the workshop, intelligent agents and web services are two important areas of my current research.

What I like about this approach is the use of two independent layers that interdependent. For example, an external public layer that provides ontologies and interfaces and an internal layer that has the intelligence for control and reasoning. This enables the construction of domain models that not only can provide code generation, but also serve as a run-time artifact. This means that these models can be interfaced with other applications as part of the Semantic Web vision and can promote both reuse and inter-operability. Thus, as I continue to develop my ideas in the domain of fMRI with a paper by Nakai et. al. (2008) at http://ipal.i2r.a-star.edu.sg/doc/publications/nakai2008mrms.pdf on "Ontology for fMRI as a Biomedical Informatics Method", I think about how to interface my domain knowledge into this domain quickly and efficiently. In their paper, Table 3 "Initial Proposal of functional magnetic resonance imaging (fMRI) ontology" they provide a beginning point for the development of a Statistical Analysis ontology as a subclass to their Parameter Description class. The statistical analysis subclass would include:

-Statistical Method
-Threshold
-Post-hoc
-Covariance

This serves a useful starting point to define the subclasses for not only the fMRI neuroimaging modality, but other modalities as well. Protégé is excellent for this with its "the SLOT of the CLASS is the FACET" approach and for a programmer such as myself well-versed in the making of classes, APIs, and database schemas, it is easily extended my thoughts to the domain models of my current expertise. Furthermore, the use of web services and the service-oriented computing (SOC) concepts lends itself naturally to this context as mentioned from the previous paper. The extensibility of this idea to other neuroimaging modalities such as magnetoencephalography (MEG), electroencephalography (EEG), and near infrared spectroscopy (NIRS) is important because at increasing levels of abstraction they all need the same analytical methods that are inherently statistical in both space and time. For example, Figure 3 in the paper shows the classification of neuroimages according to these modalities and Figure 5 presents the two data flows: a) Science Flow and b) Clincial Flow. My interest is in the work of (a) through testing research hypotheses about the validity and accuracy of different statistical techniques and paradigms used in the data analysis component.

Ultimately, the choice of what technique to use will have nothing to do with availability, but everything to do with an intelligence engine based on agent technologies that will be able to recognize the characteristics of the data set and apply the correct model for the given situation. This is especially important because of the heterogenous nature of medical data sets. Also, because of the eventual creation of domain drive code at run-time, it is possible for statistical models to be created with numerical approaches that might prove to be intractable from theoretical descriptions. This is a very rich field of research with almost unlimited potential-conditional moments, i.e. mean, variance,skew, etc. of multivariate distributions of four dimensions, i.e. 3 for space and 1 for time, accounting for and correcting multi-dimensional linear or non-linear correlation in higher conditional moments. Why? Because of the needed independence assumption for feature extraction necessary to identify the outcomes of stimulus-response models. A good first step is to clarify the ontology for this subclass and how it can be represented in an architecture as part of the semantic web vision and identify what needs to be done to speed up the delivery of high-quality, applicable statistical techniques that can be placed into current commerical applications. A fun talk and presentation for the conference, and then its off to riding the canal boats, standing in front of the Night Watch, comparing my paintings to Van Gogh and listening to some jazz in Amsterdam!

References

Nakai, T., E. Bagarinao, Y. Tanaka, K. Matsuo, and D. Racoceanu (2008). "Ontology for fMRI as a Biomedical Informatics Method", Magn Reson Med Sci, Volume 7, No. 3, pp.141-155.