It is difficult to automatically detect tumors and extract lesion boundaries in ultrasound images due to the variance in
shape, the interference from speckle noise, and the low contrast between objects and background. The enhancement of
ultrasonic image becomes a significant task before performing lesion classification, which was usually done with
manual delineation of the tumor boundaries in the previous works. In this study, a linear support vector machine (SVM)
based algorithm is proposed for ultrasound breast image training and classification. Then a disk expansion algorithm is
applied for automatically detecting lesions boundary. A set of sub-images including smooth and irregular boundaries in
tumor objects and those in speckle-noised background are trained by the SVM algorithm to produce an optimal
classification function. Based on this classification model, each pixel within an ultrasound image is classified into either
object or background oriented pixel. This enhanced binary image can highlight the object and suppress the speckle
noise; and it can be regarded as degraded paint character (DPC) image containing closure noise, which is well known in
perceptual organization of psychology. An effective scheme of removing closure noise using iterative disk expansion
method has been successfully demonstrated in our previous works. The boundary detection of ultrasonic breast lesions
can be further equivalent to the removal of speckle noise. By applying the disk expansion method to the binary image,
we can obtain a significant radius-based image where the radius for each pixel represents the corresponding disk
covering the specific object information. Finally, a signal transmission process is used for searching the complete breast
lesion region and thus the desired lesion boundary can be effectively and automatically determined. Our algorithm can
be performed iteratively until all desired objects are detected. Simulations and clinical images were introduced to
evaluate the performance of our approach. Several types of cysts with different contours and contrast resolutions images
were simulated with speckle characteristics. Four thousand sub-images of tumor objects and speckle-noised background
were used for SVM training. Comparison with conventional algorithms such as active contouring, the proposed
algorithm does not need to position any initial seed point within the lesion and is able to detect simultaneously multiple
irregular shape lesions in a single image, thus it can be regarded as a fully automatic process. The results show that the
mean normalized true positive area overlap between true contour and contour obtained by the proposed approach is
90%.
|