Seed Point

The first step of the method focuses on finding the starting point (or seed point) from which the contour will start growing. Special care has to be taken on estimating this point which directly affects the accuracy of the segmentation. As stated before, mammographic image segmentation presents difficulties mainly due to the low contrast in the skin-line and to the non-homogeneous background. From our experience this lower contrast is less severe for points close the nipple. Therefore a seed point can be easily detected in points at this area. An initial guess of a seed point is obtained as the first local maxima of the gradient in the scale space representation along the $x$ axis at half the height of the image. Obviously, this first estimation lacks of robustness if this first local maxima does not correspond to the skin-line. That could be the case if the point lies inside the breast area (due to a low contrast of the skin-line) or in the background (due to noise, label and other image artifacts). A more robust approach is adopted based on analyzing the position of various seed points at close the same position (at a small range in the $y$ coordinate). The final seed point is obtained using a least median error estimation. Edge direction will also provide an important information in the contour growing process. Therefore the estimation of the initial angle it is also important. In this case a similar least median error estimation is adopted for the angle measure. Figure (a) shows an example of seed detection.