Structuring Elements in the Watershed Algorithm for the Segmentation of Mammography Images

Abstract

Reading mammography images has always been a challenging task even for experienced radiologists. With the advancements in computer technology, machine tools such as the Computer Aided Detection and Diagnosis (CAD) systems are widely engaged as a second reader to assist radiologists in image reading. One of the important processes in the CAD machine is the segmentation process. The morphological watershed algorithm is one of the hybrid technique that combines boundary and region criteria, but this algorithm has several drawbacks such as over-segmentation and sensitive to noise. In this research, the denoising method applies the Principal Component Analysis (PCA) filtering. Prior to the segmentation by the watershed algorithm, the Fuzzy C-Means (FCM) clustering algorithm is used to identify the image foreground, which is the region of interest (abnormality region). A marker-controlled watershed algorithm is implemented to overcome the over-segmentation drawback. Furthermore, applying a suitable shape of structuring element in the watershed algorithm has the same effect of reducing the over-segmentation problem. Thus, three shapes of structuring elements, which are the disk, diamond, and octagon are tested and compared. The aim of this research is to find a suitable shape of structuring element for the marker-controlled watershed algorithm. For the evaluation of the segmentation performance, three evaluation methods are used, which are the Jaccard Index (JI), Dice Similarity Coefficient (DSC) and Figure of Merit (FOM). The result of the comparison shows that the diamond-shaped structuring element is a suitable shape for the segmentation of mammography images.