Shape analysis for classification of breast nodules on digital ultrasound images

Abstract

One of the imaging modalities for early detection of breast cancer malignancy is ultrasonography (USG).  The malignancy can be analysed from the characteristic of nodule shape.  This study aims to develop a method for classifying the shape of breast nodule into two classes, namely regular and irregular classes.  The input image is pre-processed by using the combination of adaptive median filter and speckle reduction bilateral filtering (SRBF) to reduce speckle noises and to eliminate the image label.  Afterwards, the filtered image is segmented based on active contour followed by feature extraction process.  Nine extracted features, i.e. roundness, slimness and seven features of invariant moments, are used to classify nodule shape using multi-layer perceptron (MLP).  The performance of the proposed method is evaluated using 105 breast nodule images which comprise of 57 regular and 48 irregular nodule images.  The results of classification process achieve the level of accuracy, sensitivity and specificity at 96.20%, 97.90% and 94.70%, respectively.  These results indicate that the proposed method successfully classifies the breast nodule images based on shape analysis.