Ambient Diagnostics - Ambient Intelligence for Scientific Discovery - page 226

or laptop computer for clinic or family use. Third, it's a unique exploration
to combine Traditional Chinese Medicine (TCM) with contemporary computer
vision technologies. The results of this project may inspire future long-term
development of bio-computing technologies and further the use of computers in
the medical field.
Studies show that there are correlations between the digestic diseases and
tongue feature changes. Here we focus on colon polyps that would likely be-
come colon cancer. This pilot study phase focuses on a preliminary investigation
of the computer-based tongue inspection technology. A set of computer vision
models have been developed to simulate the TCM diagnosis, for example, images
that showed detectable cancer signatures such as color and coating texture of
the tongue. Those visual feature descriptions will eventually be integrated into
a decision-making model that will help to generate the final diagnosis or con-
clusion: normal, abnormal, likelihood of cancer, etc. The software includes the
following issues: 1) Segmentation: The raw images are preprocessed with color
normalized so that they have better numerical representation. Then each image
is segmented to remove the background. The Deformable Template algorithm
is applied to generate an accurate outline of the tongue. After the initial image
processing, the improvement of the color normalization and segmentation is in-
vestigated. 2) Texture feature extraction: The texture, which includes cracks and
distributions of the tongue proper, is the most important feature. It is the most
challenging task in the project because the texture is not uniformly distributed
and the orientation and size varies from image to image. 3) Visualization models:
With the color measurement and texture features, a set of visualization methods
are explored. 4) Diagnosis with Neural Computing: Artificial neural networks are
used to classify samples.
3 Tongue Imaging
We have explored two scientific methods for tongue imaging so that we can
recover the realistic measurement of physical values. The first approach is to use
a modified hand-held color scanner with a microscopy slide on top of the tongue.
As the scanner is gently moved from the root of the tongue to the tip, a flat
image can be obtained. Figure 1 shows a sample image. The advantage of this
method is its simplicity; it can avoid major color calibration and the removal of
artifacts. However, it is a contact measurement that we want to try to avoid in a
clinical environment, and the hardware needs to be specially designed to fit the
size of tongues. The second approach is to take a picture of the tongue with a
commercial digital camera (640 x 480 pixels) plus a Munsell ColorChecker [27]
embedded inside the image. Since we already know the color value of the test
cells on ColorChecker, we can calibrate the color of the image computationally.
Because the color in an image varies with cameras, lighting and equipment
settings, we had to calibrate the color for each image before the analysis. We
used the Mansall color calibration board and the newly developed color cali-
bration software. Before the camera took a tongue image, the operator took a