Automatic mass segmentation method in mammograms based on improved VFC snake model - Emerging Trends in Image Processing, Computer Vision, and Pattern Recognition - page 171

3.2.2 Segmentation accuracy
To further explain the accuracy of our algorithm, we introduce another area-based evaluation
method then. The area overlap ratio criteria are the most common evaluation criteria in medic-
al images, which are the ratio of overlapped area between the segmented region of VFC Snake
method and the criterion region of ground truth segmented manually. The performance of the
proposed method and the typical VFC Snake method is tested on the successfully detected
mammograms of MIAS by the following equation:

(15)

where L is the area segmented by VFC Snake model, T is the area of ground truth. S L ∩ T and
S L ∪ T are the intersection area and union set area of the two regions, respectively. The average
area overlap ratio and the variance of the segmentation results are shown in Table 2 . We can
see that the average area overlap ratio of improved method is much higher than the typical
method, and the variance is much lower, that is to say, our auto-segmented results are gen-
erally much more close to the ground truth. It is proved that our approach indeed performs
much more excellent results compared with the typical method.

Table 2
Area Overlap Ratio of Different Methods

Method Mean (%) Variance (%)

Typical method 76.0151 11.5249

Improved method 90.4073 2.1556

3.2.3 Segmentation similarity
Finally, we introduce a new measure method of medical image segmentation which is based
on boundary distance-based similarity [ 20 ] . The above-mentioned area-based criteria have re-
lected the difference between the region of arithmetic-segmented result and manual-segmen-
ted result, while it could neither reflect the difference nearby the contour nor estimate whether
the arithmetic-segmented curve is bigger or smaller than the ground truth.
The new evaluation index is described as expression (16) , where m outside and σ outside are the
mean and standard deviation of the bigger portion, which consist of points from segmented
contour outside the ground truth (segmented manually), and m inside and σ inside are the mean
and standard deviation values of the smaller portion, which are inside. R equ is the radius of a
circle equivalent to the segmented region. Here, the value of σ reflects the level of similarity,
the smaller the σ value is, the higher similarity level has been reflected.