Robust detection of abnormality in highly corrupted medical images

Abstract

Medical imaging helps to detect and monitor internal irregularities in the human body. We leverage a block median filtering technique to model pixel-to-pixel differences between two images to develop automated detection of abnormalities in noisy medical images. We propose two robust detection methods, with the test statistic being the conventional maxima and the scale-invariant ratio of the medians from partitioned image grids. Theoretically, we investigate the asymptotic behaviors of two proposed tests. Numerically, we carry out simulation studies to investigate the type I error rate and the power of two tests. In addition, a real application in medical images with gastrointestinal bleeding demonstrates the outperformance and efficiency of the ratio test method. Besides, the developed tests can also be applied to problems in other scientific fields, e.g., air pollution detection using collected remote sensing hyperspectral images.