Kernel Size Dependent Outlier Separation in Images

The ability of a novel filtering technique to separate outliers of different scales is demonstrated in a systematic way. The filter utilizes a kernel level modified Z distribution to find outliers separating the input image into a filtered image and an outlier image at the end of every iteration. The length-scale of outliers extracted in the outlier image at every iterarion is shown to depend on the kernel size used in the filtering process. The outlier image, formed of the outlers flagged by the filtering process is shown to contain valuable information about the input image. The ability to detect outliers at different length scales even under low contrast scenarios can be significant in discovering early stage calcifications and nodules as outliers in tissues. The possibility of such discoveries is demonstrated by showing that the filter successfully separates artifacts of certain length scales when applied to an ultrasound image of a thyroid tissue sample iteratively.