Ambient Diagnostics - Ambient Intelligence for Scientific Discovery - page 227

Fig. 1. First image from the tongue scanner.

picture of the calibration board and saved it along with the tongue image. Then
the computer performed the color calibration for the data. At this phase, we
developed and tested the color calibration software.
A semi-automatic color calibration tool was developed for the project. By
manually clicking the four corners of the ColorChecker, the software can per-
form the transformation and find the points in each square. Then a linear color
calibration model is used to recover the original color of the tongue under various
lighting conditions [27].
Parallel to the data collection from cancerous patients, tongue images were
also collected from healthy subjects for studying the range of the deviation and
mean of the 'normal' tongue images, such as RGB color space, coating texture,
etc. Over 17 'normal' tongue images were collected in the database. Those images
helped to establish a baseline for a 'normal tongue.'
We tested 17 tongue images taken from a “healthy” individual with a digital
camera under different combinations of illumination and lighting orientations.
The illumination conditions were daylight and indoor lighting in an oce. 160
points from each tongue image were sampled to generate the data as shown in
Table 1.

4 Segmentation of Tongue Image

There are many ways to segment the tongue area from the background. Color-
based segmentation is the least reliable way because of the variations of tongue
color and shadows. Active Contour may overcome the color variation problems
by tracking the gradient of the intensity along the tongue edge. The typical al-

Table 1. Color variations of a normal tongue under different conditions.

Color Before Calibrated After Calibrated
Space Mean STD Mean STD

R 0.4604 0.0439 0.6135 0.0339

G 0.4141 0.0323 0.4940 0.0288

B 0.4632 0.0492 0.5066 0.0288