Local Segmentation by Large Scale Hypothesis Testing: Segmentation as outlier detection

We propose a novel and efficient way of performing local image segmentation. For many applications a threshold of pixel intensities is sufficient. However, determining the appropriate threshold value poses a challenge. In cases with large global intensity variation the threshold value has to be adapted locally. We propose a method based on large scale hypothesis testing with a consistent method for selecting an appropriate threshold for the given data. By estimating the prominent distribution we characterize the segment of interest as a set of outliers or the distribution it self. Thus, we can calculate a probability based on the estimated densities of outliers actually being outliers using the false discovery rate (FDR). Because the method relies on local information it is very robust to changes in lighting conditions and shadowing effects. The method is applied to endoscopic images of small particles submerged in fluid captured through a microscope and we show how the method can handle transparent particles with significant glare point. The method generalizes to other problems. This is illustrated by applying the method to camera calibration images and MRI of the midsagittal plane for gray and white matter separation and segmentation of the corpus callosum. Comparing this segmentation method with manual corpus callosum segmentation an average dice score of 0.88 is obtained across 40 images.
Keyword: Segmentation, outlier detection, large scale hypothesis testing, locally adjusted threshold