NAMIC Wiki:ShapeAnalysis:AHHFeb05SLC Minutes

Minutes from the Shape Analysis Discussion at the AHM in Salt Lake City, Feb 2005

We had a quite lively and productive discussion with clinical and methodological researchers regarding the goals of the shape analysis community within NAMIC (and outside of it). The main topics that were discussed include the following:

• Methodological issues in regard to shape representation: What is the meaning of 'shape'? There are a series of shape representations in NAMIC, but the definition of shape is still not clear and not fully understood. This is especially true for the shape of groups of objects and not just single objects.
• Classification vs Regression/Correlates: There is a clear need on the clinical side for future shape analysis methods to yield correlates/covariates to other subject-related features such as clinical variables, test scores, genetical background etc.
  • Classification is important, but needs to be corrected for a series of covariates, such as gender, age, onset of illness, drug type etc. Current shape analysis methods can only use discrete groups. Methods for correction or studying these covariates do not yet exist due to the non-linear or non-Euclidean nature of many shape features. Approximatively linear regression can be employed but is not fully correct for these shape features. Also, issues of local vs regional vs global correction are not yet solved.
  • Clinicians would like to correlate hippocampal shape with different genetical alleles, test scores, or specifically in schizophrenia with negative and positive symptoms.
• Need for standardization framework for conducting population based studies, not just shape based studies. Core 2 involvement will be instrumental here (see below).
• Results of statistical methods need to be visualized to the clinicians as much 'intuitive' as possible. Need for visualization of not just the significance maps, but effect size maps, metric maps, covariance maps. A clear description of how to interprete these maps should be available to the clinicians.
• Many participants mentioned the need for studying the differences between different representation, normalization methods and statistical methods theorectically and empirically is very relevant. NAMIC with its variety of methods has a chance to do this based on the above mentioned framework.
• Many people stressed the importance of statistical tests and lack of a satisfactory solution on how to derive statistical significance of the detected patters for high dimensional data, such as shape descriptors or other image data. Multiple comparison corrections, false-discovery rate, permutation tests were mentioned as options, but it was generally agreed that we need to spend more effort in developing methods that will evaluate the validity of the detected differences.

Framework for population studies

In the course of the discussion it became clear that more than just a simple solution for a standardized framework within NAMIC for shape analysis studies was needed. There was general agreement that the shape analysis framework should be generalized as much as possible to any statistical study of populations, such as DTI fiber properties, DTI images or even fMRI studies. This framework should be able to handle

• different kinds of 'FeatureSet' objects, e.g. SPHARM or PDM surfaces, skeletons, L2 basis descriptors
• statistical methods
• alignment methods
• correspondence methods

Other observations:

• Core 2 should definitely be involved in the production of this framework
• The NAMIC shape analysis community currently has a first draft for describing populations and the feature objects. This is basically the input and the result of the individual representation 'building' step for such studies.
• No draft for a framework for
  • Normalizing reprepsentations, i.e. alignment methods, scaling methods, correspondence establishing methods
  • Application of statistical methods such as statistical comparisons, tests, model building
• the output of statisical methods can be again a feature set (e.g. from a PCA model building)
• The results of this frameworks should be visually interpretative including not just significance maps, but also 'QC' maps such as effect size maps, covariance maps.