Automatic mass segmentation method in mammograms based on improved VFC snake model - Emerging Trends in Image Processing, Computer Vision, and Pattern Recognition - page 170

and with stronger capability of convergence than the typical method, that is, our approach
performs much beter when we segment the masses in mammograms from DDSM database.
The MIAS database has offered the central coordinate and radius of each abnormal region
showing in Figure 6 ( a-3)-(d-3), as well was the ground truth segmented manually. Here, we
also initialize the VFC Snake model using the parametric circle obtained by mass location.
From Figure 6 ( a-2)-(d-2), we observe that the results segmented by the typical method exist
serious distortions and the contours can hardly converge to the real boundaries. Compared
with the typical model, our proposed method can completely remove the labels or interference
and achieve more robust and accurate results. As we can see, the curves are much more close
to the ground truth and precisely tend to the object even in blurry regions.
From the enlarged results of our method, we find that the margin of the last image is rough
and the other ones are smooth, that is because the severity between the last lesion and the rest
are different, the last lesion is malignant while others are benign. Our results objectively re-
lect the pathology characteristics of actual masses to some extent that the malignant masses
are always with burrs. This performance is somewhat benefit to the early diagnosis of breast
cancer. Therefore, for a CAD system, we are able to extract the features of our detected results
and determine the severity of abnormalities for a further work to give a considerable “second
suggestion” to the clinician.

3.2 Algorithm Performance Analysis
We test the proposed method on the DDSM and MIAS database and evaluate the performance
from three aspects.

3.2.1 Detection rate
First, we compute the detection rate, our evaluation principle is that the autosegmented region
by the proposed method is completely within the criterion region by the experts. In the case of
DDSM database, the criterion region is the outline formed by chain code data, and for MIAS
database, the criterion region is the circle formed by the center coordinates and the radius. The
detection rates of masses for each database are shown in Table 1 . As we can see, 362 images
are successfully extracted in total and the average detection rate is 90.5%, and even reaches up
to 91.47% for the DDSM images. While it is lower for the MIAS images, the lesions in dense
breast images of MIAS are always embedded in the gland and we can hardly obtain the mass
position by location or edge map by edge detection operator for the deformable model.

Table 1
Mass Detection Rate by the Proposed Method

Database Tested Images Detected Images Nondetected Images Detection Rate (%)

DDSM 340 311 29 91.47

MIAS 60 51 9 85.0

Total 400 362 38 90.5