A simulator for evaluating methods for the detection of lesion-deficit associations

Abstract

Although much has been learned about the functional organization of the human brain through lesion-deficit analysis, the variety of statistical and image-processing methods developed for this purpose precludes a closed-form analysis of the statistical power of these systems. Therefore, we developed a lesion-deficit simulator (LDS), which generates artificial subjects, each of which consists of a set of functional deficits, and a brain image with lesions; the deficits and lesions conform to predefined distributions. We used probability distributions to model the number, sizes, and spatial distribution of lesions, to model the structure-function associations, and to model registration error. We used the LDS to evaluate, as examples, the effects of the complexities and strengths of lesion-deficit associations, and of registration error, on the power of lesion-deficit analysis. We measured the numbers of recovered associations from these simulated data, as a function of the number of subjects analyzed, the strengths and number of associations in the statistical model, the number of structures associated with a particular function, and the prior probabilities of structures being abnormal. The number of subjects required to recover the simulated lesion-deficit associations was found to have an inverse relationship to the strength of associations, and to the smallest probability in the structure-function model. The number of structures associated with a particular function (i.e., the complexity of associations) had a much greater effect on the performance of the analysis method than did the total number of associations. We also found that registration error of 5 mm or less reduces the number of associations discovered by approximately 13% compared to perfect registration. The LDS provides a flexible framework for evaluating many aspects of lesion-deficit analysis.

Subjective Probability

“…The proposed classification framework was experimentally evaluated on synthetic data (mixtures of Gaussian distributions), on realistic brain lesiondeficit data generated by a simulator [33] conforming to a clinical study [44], and on real fMRI brain activation distributions obtained from a study that explores neuroanatomical correlates of semantic processing in Alzheimer's disease [45]. These datasets as well as the experimental results are described below.…”

“…For this purpose we estimate the mean and covariance matrix of data, or apply the expectation-maximization (EM) algorithm [27,29] and its variant, the k-means algorithm [33]. (2) Given a 3D image belonging to an unknown sample, characterize the corresponding distribution of ROIs.…”

“…2, avoids the costly computation of probabilities P ij in the E step by assigning each data example z i to a Gaussian component, whose mean m j is the closest to z i based on the Euclidean distance (see Eq. (A.8) in Appendix A for definition) [33]. This is equivalent to the partitioning of a dataset S (containing the examples z i ) into k clusters.…”

“…We performed classification of realistic brain lesion distributions (available at http://denlab.temple.edu/data_repository) that were generated using a lesion-deficit simulator (LDS) [33] with the spatial statistical model conforming to the Frontal Lobe Injury in Childhood (FLIC) study [44]. In implementing the LDS we used probability distributions to model the number, size, and spatial distribution of lesions, as well as registration error and structurefunction associations.…”

“…From these parameters, the simulator generates a complete dataset, including spatial lesions and clinical variables. The LDS has been successfully used in evaluation and scalability testing of data mining techniques and to provide an approximate number of subjects needed to discover certain associations so that medical experiments can be planned accordingly [33]. The segmentation of ROIs in the FLIC study was performed manually by a neuroradiologist using thresholding.…”