Ambient Diagnostics - Ambient Intelligence for Scientific Discovery - page 235

is to scan the image and keep track of how often pixels that differ by ∆z in
value are separated by a particular distance d in position [49, 41]. Based on the
co-occurrence matrix, we compute the energy function and entropy function for
further texture descriptions. The energy function, or angular second moment,
is an image homogeneity measure; the more homogeneous the image, the larger
the value.

H entropy = − ( P [ i, j ]log 2 ( P [ i, j ])) (9)

i,j

The entropy function, can also be used as a measure for “textureness or
complexity.”

P [ i, j ]
1+

H entropy = − (10)

|i − j|

i,j

where, P is a gray level co-occurrence matrix that contains information about
the position of pixels having similar gray level values.

6 Visualization Techniques and Process

In this section, we show how visualization can provide insights and facilitate the
analysis and clustering of tongues based on feature values. We have a feature
vector: F=[a*,b*,Db,CI,energy,entropy]. Li & Cai extracted the values for these
six features for a set of 34 tongues which belong to people of five different diag-
nostic categories: Healthy (H), History of Cancers (HC), History of Polyps (HP),
Polyps (P), Colon Cancer (C) [50]. We used this data set to demonstrate the
visualization techniques and process.

6.1 Cluster Plots

The goal of data exploration is to investigate if there are some obvious corre-
lations between different features. This can be achieved by 2D and 3D cluster
plots. We found that there are definite patterns for some categories as summa-
rized in Table 2. Figure 6 shows a 2D cluster plot of Db-entropy which shows a
clear pattern for Healthy (H) and Polyps (P) cases.
As a* and b* are two chromatic dimensions in the L*a*b* color space, ob-
served definite ranges of values for these features for certain diagnostic conditions
indicate that there is a strong correlation between the tongue color and these
diagnostic conditions. We also observe two outliers with very high crack index,
one with a Healthy condition and one with History of Polyps and History of
Cancer. This suggests that these are special cases where these tongues normally
have lots of cracks, and the amount of cracks therefore does not reflect on the
disease condition.
For 3D cluster plots, we used different colors and glyphs for each category (see
Fig. 7) and allowed users to rotate them around each axis to facilitate viewing.
Rotating the plots increased the sense of depth and the perception of clusters
and correlations. We also observed that the high precision of the raw data might