Other model based approaches include deformable

Other model-based approaches include deformable models such as active contours and level sets. Snake active contours have been applied to BUS segmentation with good results. In Jumaat et al. (2011), parametric active contour models such as gradient vector flow and balloon were used in BUS mass segmentation, after pre-processing with median filtering and histogram equalization. A segmentation refinement stage was designed, integrating curvature information or even empirical knowledge to improve the initial result. Another method described by Madabhushi and Metaxas (2003) relies on the automatic definition of seed points based on empirical knowledge given by radiologists. Region growing was then applied to obtain an initial contour. Image and texture information was used to classify the pixels, and the boundary points found with a directional gradient served as an initial contour for an active contour model, which used the directional gradient as a stopping criterion.
A level set model was applied to the segmentation of lesions in BUS (Huang et al. 2007), yielding better results when compared with active contours. The initial contour was obtained through binary thresholding after use of the modified curvature tgf beta receptor equation (MCDE) to remove noise and enhance the image contours. In Chang et al. (2005), the ultrasound image was first processed using anisotropic diffusion filtering. Then, an initial contour for the level set segmentation was obtained combining the stick method with binary thresholding. Despite the good segmentation results and low noise sensitivity, segmentation using deformable models, such as active contours or level set models, has a number of significant drawbacks with respect to the automatic definition of an appropriate initial contour as well as the high sensitivity to local minima. Also, the convergence of the active deformation process may be computationally heavy.
In Chang et al. (2010), a modified watershed algorithm was used for semi-automatic contour extraction of BUS lesions. Morphologic operations were first applied to obtain more accurate contours and prevent typical watershed oversegmentation (Sarpe 2010). The watershed transform was also used in Zhang et al. (2011) for automatic lesion detection. Images were pre-processed by mean filtering and fuzzy logic histogram thresholding. In both cases, watershed segmentation achieved high accuracy, similar to that of manual segmentation.
A k-means algorithm was tested in Boukerroui et al. (1998) to achieve the segmentation of BUS images. The authors classified tissues using an adaptive method based on texture information. Even though the algorithm is simple and effective, the results were dependent on system parameters and a major problem might have arisen from eventual similarities between the mass and shadows or other artifacts in the image, which could lead to inclusion in the wrong cluster of pixels.
Classification algorithms have been widely considered as segmentation alternatives for a number of image modalities. In the specific case of BUS, several studies have used support vector machines (SVMs) and neural networks (NNs) to obtain lesion contours. In Chen et al. (2002), the authors successfully combined wavelet analysis of the image with an error retro-propagation NN, using contrast variance and autocorrelation as inputs. Another study (Huang and Chen 2004) classified BUS images employing a watershed segmentation algorithm along with a NN trained using texture descriptors. A Bayesian NN with five hidden layers was also tested in BUS image segmentation in Drukker et al. (2002). Texture, gradient and acoustic information of the images was retrieved to train the NN, which was used to validate candidate regions obtained with a region-growing algorithm, starting from points of interest defined using the image gradient. However, the method proved to be unreliable, especially where lesions were not uniform. In Shan et al. (2012), a similar seed point approach was used, along with pixel classification using NN. Multidomain features such as intensity, texture, phase in max-energy orientation and radial distance were combined. This method yielded interesting segmentation results in BUS. Su et al. (2011) applied self-organizing maps, using textural local information, to obtain an initial contour. This outline would later be segmented with active contours, culminating in a fully automated method with good accuracy. Combining SVMs with textural information, Liu et al. (2010) proposed a robust, high-precision method for mass segmentation in BUS. Although the use of classifiers for targeting ultrasound images has had promising results, the training required and the selection of an appropriate set of features for its application can make the task complicated and time consuming (Huang and Chen 2004).