When Less is More: Improvements in Medical Image Segmentation through Spatial Sub-Sampling

Segmentation is a common task in medical image analysis. It is frequently solved by fitting an intensity model, consisting of distributions for each pure tissue and each partial volume tissue combination, to the intensity histogram of the image data. However, this approach discards any spatial information present in the data. We present a method that recovers some of this information via regional sub-sampling during the fitting process. Experiments are performed on simple simulated data, simulated MR images from Brainweb, and real MR data from eight young normal subjects. The spatial sub-sampling procedure is shown to significantly improve the segmentation stability.