Initial lesion detection and region of interest labeling in ultrasound breast images

Abstract

The proposed work is focused on initial lesion detection and region of interest (ROI) labeling in ultrasound breast images. The aim is to automate the manual process of ROI labeling by radiologists in computer-aided diagnosis (CAD). In our research, we propose a statistical method, namely local mean, in initial lesions detection and investigate the effectiveness of using k-means clustering in automated ROI labeling. A total of 380 ultrasound breast images have been used as samples. Histogram equalization is used to pre-process the images followed by hybrid filtering and marker-controlled watershed segmentation. The minimum local mean of the identified segments is then used to determine the initial lesion. Subsequently k-means clustering is used to identify segments similar to the initial cluster, which are finally combined to form the ROI. The accuracy of the automated ROI labeling is measured by an overlap of 0.4 with the lesion outline compared to the lesions labeled by the radiologist. We compare the proposed method with the popular radial gradient index filtering technique, and demonstrate its improved performance. We conclude that the proposed method is accurate in automated ROI labeling of cysts and malignant lesions, but not for fibroadenoma.