Looking into Algorithms to Evolve a Robust Classification of Ultrasound Image of Human Liver Steatosis Using Two Independent Image Sources

Abstract

Supervised classification is found to be data specific. A single data set is usually partitioned into training and test subsets usually with several fold cross validation. An ideal classification model trained with exhaustive variation in training examples should work with other similar but unseen data from an independent source. Present authors developed an efficient algorithm to classify fatty liver versus normal liver from ultrasonogram (US) of human liver. In this article it is explored how far that algorithm is applicable for classification of US image obtained from two different sources and how the classification accuracy is affected if two data sets are pooled together. Finally we explore the possibility of training by one source and testing by another source. Idea is to find out a robust algorithm which works in a portable environment.