Ambient Diagnostics - Ambient Intelligence for Scientific Discovery - page 241

Table 5. Test samples (not included in the training dataset).

data set target Db CI a* b* energy entropy

1 10 (P) 1.68 0.62 18.76 9.97 0.25 4.76

2 10 (P) 1.64 0.55 9.98 5.35 0.33 4.66

3 10 (P) 1.73 1.92 7.75 4.73 0.33 4.59

4 1(NP) 1.63 0.1 11.55 8.77 0.33 4.91

5 1(NP) 1.61 0.61 24.56 6.08 0.16 4.64

6 1(NP) 1.72 1.17 11.53 11.34 0.26 4.46

Table 6. Results for two types (spread factor = 0.4).

data set target GRNN PNN

1 10 (P) 10 (P) 10 (P)

2 10 (P) 9.0049 (P) 10 (P)

3 10 (P) 1(NP) 1(NP)

4 1(NP) 1(NP) 1(NP)

5 1(NP) 1(NP) 1(NP)

6 1(NP) 1(NP) 1(NP)

would work well with a larger sample size. There is still a long way to go before
we can come up with a selective and robust classification method for the tongue
inspection.

8 Conclusions

Ambient Diagnostics is a contemporary technology that is inspired by ancient
medical practices. The goal is to detect abnormities from seemly disconnected
ambient data. In this chapter, we focused on computerized tongue inspection.
Our research started with collecting tongue samples at a clinical lab setting
and building conceptual prototypes for scientific discovery along the way. The
explorations include digital imaging, color calibration, feature descriptions, vi-
sualization and neural computing. From this preliminary study, we have learned
the following lessons:
The portable tongue scanner is more reliable than the digital camera in terms
of invariance of illumination, reflection and angles. However, its resolution is low
and cost is rather high. We will do further investigation to reduce the cost and
increase the resolution.
Previous TCM studies have shown strong correlations between the color of
tongue coating and cancers. We found that the texture characteristics on the
tongue surface is more sensitive to the colon polyps (pre-cancerous) or history of
polyps than color characteristics. This discovery will lead us in a new direction
toward effective tongue feature expressions, such as adding more texture de-
scribers in the feature vector. The more dimensions of the describers, the more
accurate the classification and recognition of the model. We had four dimensions: