Elsevier

NeuroImage

Volume 47, Supplement 1, July 2009, Page S102
NeuroImage

Spatial Clustering of Response Curves

https://doi.org/10.1016/S1053-8119(09)70885-7Get rights and content

Introduction

In many fMRI experiments the voxel-wise amplitude of the BOLD response is estimated for several different levels of a stimulus. The difference in response between levels is typically studied using an appropriately defined contrast. However, if the levels can be viewed as continuous it could be beneficial to model the relationship between the two variables (e.g. amplitude and the temperature of painful heat stimulus) as a continuous response curve observed at certain discrete points. Valuable information can potentially be gained by studying the behavior of the response across levels and voxels. In this work we develop a flexible approach for modeling and spatially clustering voxel-wise functional response curves for multi-subject fMRI data. The goal is to cluster the brain into regions with similar response curves over levels of a stimulus, and to estimate these region-wide curves and their associated variability. We apply functional data analytic modeling techniques to curves at each voxel, and employ a new model-based unsupervised spatial clustering algorithm to estimate regions with homogeneous response profiles. Our approach is general and can be applied in a variety of experimental settings. It is particularly useful when data is observed at sparse, irregularly spaced levels of a continuous predictor variable.

Section snippets

Methods

Our data consists of a pain study (n=20) comparing brain responses to noxious heat at 4 different temperatures. The temperatures differed across subjects, and were chosen so that subjects were matched on subjective pain perception. Fig. 1 illustrates the format of the data. Response to a given temperature at each voxel was defined as the peak of the BOLD signal, and was modeled using a polynomial spline basis. Coefficients were defined as parameters in a random-effects model, allowing us to

Results

Preliminary results, based on an analysis of 5 subjects using 4 clusters, are shown in Fig. 3. One cluster, located primarily in the temporal cortex, insula, thalamus, and right ventrolateral prefrontal cortex (dark red in Fig. 3), shows a mean response which increases with temperature. Each of these regions shows consistent responses to painful stimuli in previous studies. Another cluster located in the visual cortex (light blue) increases, then decreases with temperature. Two clusters show

Conclusions

We have presented a general method for modeling linear or nonlinear brain responses to a continuous predictor sampled at discrete levels. The levels need not be uniform across individual subjects. Thus, the model can be applied over a broad variety of experimental designs, and may be useful for fine-grained analysis of stimulus-response relationships. Incorporation of spatial information allows for efficient estimation of contiguous regions with similar response profiles.

References (0)

Cited by (0)

View full text