Computerized lesion segmentation of breast ultrasound based on marker-controlled watershed transformation

This paper presents a computerized segmentation method for breast lesions on ultrasound (US) images.

It consists of first applying a contrast-enhanced approach, i.e., a contrast-limited adaptive histogram equalization. Then, aiming at removing speckle and enhancing the lesion boundary, an anisotropic diffusion filter, guided by texture descriptors derived from a set of Gabor filters, is applied. To eliminate the distant pixels that do not belong to the tumor, the resulting filtered image is multiplied by a constraint Gaussian function. By doing so, both the segmentation and the marker functions are generated and could be used in the marker-controlled watershed transformation algorithm to create potential lesion boundaries. Finally, to determine the lesion contour, the average radial derivative function is evaluated. The proposed method was tested with 50 breast US images and 60 simulated "ultrasound-like" images. Accuracy and precision of the segmentation method were then assessed. For the accuracy, three parameters were used: Overlap ratio (OR), normalized residual value (nrv), and proportional distance (PD) between contours.

The average results for US images were OR = 0.86 +/- 0.05, nrv = 0.16 +/- 0.06, and PD = 6.58 +/- 2.52%. For simulated ultrasound-like images, a better performance (OR = 0.92 +/- 0.01, nrv = 0.08 +/- 0.01, and PD = 3.20 +/- 0.53%) was achieved.

The segmentation method proposed was capable of delineating the lesion contours with high accuracy in comparison to both the radiologists' delineations and the true delineations of simulated images. Moreover, this method was also found to be robust to human-dependent parameters variations.