Edge complexity and partial volume effects.

Abstract

PURPOSE: The accuracy of MR brain image segmentation is limited by so-called partial volume effects. We hypothesized that "edge complexity" (i.e., tissue class interface border complexity) significantly influences the magnitude of such effects. METHOD: To investigate partial volume effects and provide a vehicle for validation of segmentation algorithm accuracy in brain MRI. We developed a computer simulation, the "gigabrain." The simulation is based on interpolated (supersampled) data from actual MR studies. The voxels are assigned to one of five compartments (gray matter, white matter, CSF, fat, or "background"), the compartment interfaces are "jittered" to add high frequency "signal" or "edge complexity," and the voxels are populated with appropriate values determined from human data, low pass filtered (based on the MR scanner's point spread function), and subsampled back to the sampling and voxel size of the original MR data set. RESULTS: In comparison studies with actual phantoms and human MR data, our simulation approach was able to produce images whose appearance and quantitative values were comparable with the actual data, but only when edge complexity was added to the original MR data. CONCLUSION: Edge complexity is a significant source of partial volume effects. MR simulations must include edge complexity to adequately test segmentation algorithms.