|
Abstract:
|
we propose a novel complex-valued spatial model via kernel convolution and develop a MCMC algorithm to detect brain activation at the voxel level. The implementation of this model is done in a computationally efficient way by parcellating the voxel-based image into several spatial regions also by kernel-based dimension reduction. Assuming that whether a voxel is activated or not is governed by a latent spatial random process, we show that the kernel-based model produces a more reasonable posterior probability map than the model using Gaussian processes, and hence leads to less false positives. In general, comparing to the non-spatial model, the spatial models encourage activation clustering and avoid falsely labeled isolated activated voxels. When magnitude data are used, our kernel-based model much improves sensitivity especially when signal-to-noise ratio (SNR) is high and contrast-to-noise ratio (CNR) is low. When complex-valued data are used, the model without explicit spatial structure could perform as good as spatial models. As a result, using entire complex-valued data is recommended when a sophisticated spatio-temporal model is not preferred.
|
