Classification of Multi Source Ultrasonogram Image of Steatosis

Aim of supervised classification is to obtain an algorithm that works across the data sets collected from different sources. Ideally a classifier, trained with sufficient variations in the input data, should be able to make reasonably good prediction on test data, obtained from arbitrary independent source. Medical images are sensitive to imaging devices and imaging conditions; therefore, classification of medical images is, typically restricted to a single data set where images are collected under similar conditions. In the present communication human normal and fatty liver ultrasonogram images from two different sources are used. When one source is employed for training and the other for testing, the classification accuracy is low. However, addition of a small fraction of data from the testable source to the training set drastically improves the classification accuracy indicating an algorithm to develop a classifier that works on data from different independent sources.