Feature extraction based on laws' texture energy for lesion echogenicity classification of thyroid ultrasound images

Abstract

Ultrasound image is commonly used to examine the malignancy characteristics of thyroid nodule. One of the important feature to diagnose of malignant thyroid nodule is based on echogenicity characteristic which is represented by the grey- level intensity of each nodule. Therefore, a computer-aided diagnosis (CADx) is necessary to examine the important features and classify which nodule is more likely to be a malignant or benign and should be given further treatment or not. This research proposes a scheme for classifying thyroid nodule based on texture features from histogram and laws' texture energy into four classes, i.e. anechoic, isoechoic, hypoechoic and markedly hypoechoic. Pre-processing step is conducted to enhance the ultrasound image followed by segmentation process using active contour without edge (ACWE). Feature extraction is performed by extracting 24 features and obtained 21 selected features based on information gain filter ranking. Then, these selected features are classified using multi-layer perceptron (MLP) algorithm. The classification results achieved the level of accuracy, sensitivity, specificity, PPV and NPV at 93.69%, 93.36%, 97.87%, 93.82% and 97.88%, respectively. These results indicate that the proposed scheme successfully classified the echogenicity of thyroid ultrasound nodules.