Novel Performance Metrics in the Classification of Microscopic Colonic Images
Authors: L. A. Gan Lim, R.N.G. Naguib, E.P. Dadios, J.M.C. Avila
Abstract
Abstract—Two classification metrics are proposed in this paper. The first one, called the mean relative difference confusion matrix, or MRDCM, is aimed at better quantifying the performance of a classifierthat outputs a spectrum of real numbers. Usually,the tendency is to use threshold values to distinctly put classified values into rigid categories. While this approach has proven to be effective in many studies, sometimes it is best to leave the “classification” or the interpretation of actual output values of classifiers to a human expert. The MRDCM has been conceptualized with this idea in mind. The other classification metric that is proposed in this paper is aimed at improving the aggregation of values in a conventional confusion matrix to calculate the accuracy of classification. This other proposed novel metric, called the Classification Performance Index or CPI, includes in its calculation the consideration of both the correct and wrong classifications. These proposed metrics were applied to a classification of microscopic colonic images into three categories, namely: normal, adenomatous polyp, and cancerous. The results show agreement with the conventional confusion matrix and accuracy metric plus more information.