Segmentation of Ultrasound Breast Images: Optimization of Algorithm Parameters

Segmentation of lesions in ultrasound imaging is one of the key issues in the development of Computer Aided Diagnosis systems. This paper presents a hybrid solution to the segmentation problem. A linear filter composed of a Gaussian and a Laplacian of Gaussian filter is used to smooth the image, before applying a dynamic threshold to extract a rough segmentation. In parallel, a despeckle filter based on a Cellular Automata (CA) is used to remove noise. Then, an accurate segmentation is obtained applying the GrowCut algorithm, initialized from the rough segmentation, to the CA-filtered image. The algorithm requires tuning of several parameters, which proved difficult to obtain by hand. Thus, a Genetic Algorithm has been used to find the optimal parameter set. The fitness of the algorithm has been derived from the segmentation error obtained comparing the automatic segmentation with a manual one. Results indicate that using the GA-optimized parameters, the average segmentation error decreases from 5.75% obtained by manual tuning to 1.5% with GA-optimized parameters.