Automatic mass segmentation method in mammograms based on improved VFC snake model - Emerging Trends in Image Processing, Computer Vision, and Pattern Recognition - page 169

FIGURE 5 Experimental results of mammograms from DDSM database: (a) - (d) The “thumb-
nails”. (a-1)-(d-1) Full raw images. (a-2)-(d-2) Results by our proposed method. (a-3)-(d-3)
Results by typical VFC Snake model. (a-4)-(d-4) Results by manually.
FIGURE 6 Experimental results of mammograms from MIAS database: (a)-(d) Original im-
ages. (a-1)-(d-1) Results by the proposed method. (a-2)-(d-2) Results by typical VFC Snake
model. (a-3)-(d-3) The ground truth.
3.1 Experiments Results
The suspicious regions of full raw images from DDSM were given by chain code data as
ground truth. This database also provided thumbnail images for visual browsing of each case
as shown in Figure 5(a)-(d) . The severity of these selected images is benign and the approxim-
ate abnormal regions have been marked in this database. To contrast, we illustrate the manual-
segmented results in (a-4)-(d-4). Here, we utilize the parametric circle obtained from mass loc-
ation procedure as the initialization of VFC Snake model and segment the lesion area. Results
gained by our method are shown in Figure 5 ( a-2)-(d-2), labels in these mammograms have
been removed already. From the enlarged images, we can clearly see that the contours con-
verge precisely to the real boundaries in all cases and are much similar to the manual ones.
For comparison, Figure 5 (a-3)-(d-3) states the same cases detected by the typical VFC Snake
model, unfortunately, these results are seriously influenced by the blurry tissues and can't de-
form to the objects completely, even the initial contours are very close to the actual boundar-
ies. Thus, we can say that the improved model is in lower dependence on the initial contour