Breast Ultrasound Segmentation Using Morphologic Operators and a Gaussian Function Constraint

Abstract — This work proposes a hybrid segmentation method based on Morphological Operators and on a Gaussian function constraint to delimitate the search dominium, both used for tumor segmentation purposes The Watershed opera-tor was also used to find tumor contour. The proposed method was firstly evaluated using gear simulated images (eight and sixteen teeth with different width and height), a power speckle noise of 10 % and a contrast ratio of 10:53 referred to the background. The hybrid method was also applied to 200 ultra-sound (US) breast images. For assessing the segmentation performance, similarity parameters (overlap ratio, RS, nor-malized residual mean square value, nrv, and mean distance between contours, dmed) based on the predefined shape or on the US tumor contour depicted by an expert as a gold stan-dard. The segmented contour of the gear simulated images obtained with the hybrid method better preserves the original shape. This finding was evident for both gears, which is par-ticularly important since the teeth mimic contour irregulari-ties. For the US breast images, the proposed method keeps the tumor boundaries without aggregating improper structures, resulting in RS> 50 % for 91 % of the images, nrv < 40 % for 70%, and dmed < 5 pixels for 77 % of the cases. These findings suggest that the hybrid method represents more realistically tumor contours, which is a vital diagnosis characteristic