Increasing DBM Reliability using Distribution Independent Tests and Information Fusion Techniques

In deformation based morphometry (DBM) group-wise differences in brain structure are measuredusing deformable registration and some form of statistical test. However, it is known that DBMresults are sensitive to both the registration method and statistical test used. Given the lack of anobjective model of group variation it has been difficult to determine the extent of the influenceof registration implementation or contraints on DBM analysis. In this thesis, we use registrationmethods with varying levels of theoretic similarity to study the influence of registration mechanicson DBM results. We show that because of the extent of the influence of registration mechanicson DBM results, analysis of changes should always be made with a thorough understandingof the registration method used. We also show that minor variations in registration methods canlead to large changes in DBM results. When using DBM, it would be imprudent to use only oneregistration method to draw any conclusions about the variations being studied. In order to providea more complete representation of inter-group changes, we propose a method for combiningmultiple registration methods using Dempster-Shafer evidence theory to produce belief maps ofcategorical changes between groups. We show that the Dempster-Shafer combination produces aunique and easy to interpret belief map of regional changes between and within groups without thecomplications associated with hypothesis testing.Another, often confounding, element of DBM is the parametric hypothesis test used to specifyvoxels undergoing significant change between the two groups. The accuracy and reliability ofthese tests are contingent on a number of fundamental assumptions made about the distribution ofthe data used in the tests. Many DBM studies often overlook these assumptions and fail to verifytheir validity for the data being tested. This raises many doubts about the credibility of the resultsfrom such tests. In this thesis, we propose to perform statistical analysis on DBM data using nonparametric,distribution independent hypothesis tests. With no data distributional assumptions,these tests provide both increased flexibility and reliability of DBM statistical analysis