Biomedical Engineering Reference
In-Depth Information
the surroundings, as depicted in Fig. 4d-f. Here we pay more emphasis on this
transform. First, due to the similar red colors as the tongue body, the lip areas
cannot be removed by Eq. (25), which also determines that it is better to initialize
the curve of the color GVF snake inside the tongue body area to avoid convergence
to the lip edges. Second, many tongue images captured from diseases contain white
and yellow colors of fur that are easily weakened. However, we argue that it has
no great effect, and our concern is the tongue's contour. Third, someone may note
that Eq. (15) is much similar to one of the
I
1
I
2
I
3
components proposed by Ohta
et al. [91], except that the latter is used to construct a color space.
It is not difficult to binarize the transformed images that are actually intensity
images. In binary images (Fig. 4g-i), we search for a set of tongue tip points
according to the following steps: (a) find the lowest position of the tongue body
by scanning line by line from bottom to top and define it as tip feature point
Q
;
(b) beginning with point
Q
, step left and right over
N
(here
N
=50) pixels,
respectively, and get left feature point
QL
and right feature point
QR
.
These
feature points are indicated in Figure 4g-i.
Due to the illuminant geometry, some shadows are formed between the upper
tip and tongue body in the tongue image. According to this prior knowledge,
a set of tongue root feature points can be obtained in the graylevel counterpart
of the color tongue image. At first, vertical projections of the graylevel images
are performed, and we can find the valley in the formed histogram. Then Sobel
operators are used to detect the edge of the tongue root. Searching near the valley
pixels, we can obtain tongue root feature point
P
. Finally, the intersecting points
of the tongue root and upper lip, i.e., left feature point
PL
and right feature point
PR
at the root of the tongue, can be obtained by boundary tracing techniques.
We now have several feature points at the tip and root boundary of the tongue.
A close curve can be formed by linking
PL
and
QL
,
PR
and
QR
, together with
the curves
PPL
,
PPR
,
QQR
, and
QQR
. These close curves (the white lines
in Figures 4p-r), are used as the initial curves of the color GVF snakes in our work.
The use of prior knowledge makes separation of the tongue body from lower
lip the main task. Before deforming the initial curves, Di Zenzo color gradients are
computed in the original images and GVF vectors are solved on the basis of them.
Figure 4m-o depicts the closeups of the local GVF force fields in the three tongue
images. Because the initial curves are relatively close to the true boundaries of
the tongue bodies, they are driven by the GVF external forces to converge quickly.
Figure 4s-u shows the final segmentation results of the tongue images.
4.1. SnakePit-Based InteractiveRefinement of SegmentationResults
Although these results are promising, the snake curve will sometime fall into
unexpected areas. For instance, Figure 5 gives two examples of failure of tongue
image segmentation by the color GVF snake. To provide correct segmentation
