Application of a neural network and four statistical classifiers in characterizing small focal liver lesions on CT

Abstract

Differential diagnosis of hypodense liver lesions on CT is a common radiological problem. The aim of this study was to apply image analysis methods on non-enhanced CT images for discriminating small hemangiomas, the most common non-cystic benign lesion, from metastases, which represent the vast majority of malignant hepatic lesions. Twenty textural features were calculated from the CT density matrix of 20 hemangiomas and 36 liver metastases and were used to train a multilayer perceptron neural network classifier and four statistical classifiers. The neural network exhibited the highest classification accuracy (83.9%) employing 3 textural features (kurtosis, angular second moment, and inverse difference moment), 2 hidden layers and 4 hidden layer nodes. The diagnostic accuracy of CT in characterizing small hypodense liver lesions may be improved by the application of image analysis methods employing a multilayer neural network classifier.