Biomedical Engineering Reference
In-Depth Information
and informative three-class classification of normal vs. adenocarcinoma vs. squamous cell
carcinoma tumor types, it is more difficult to achieve accurate classification than for the
two-class system.
In Figure 1.52, we present a representative RadViz display of the classification output
for the three-class problem. Here, only 15 genes most effective at tissue classification were
selected from the entire starting population by the statistical algorithm underlying the
RadViz tool (177). This RadViz display illuminates the overall classification results nicely
and in a visually intuitive manner. In this display, the specific genes located on the partic-
ular labeled circumference regions are those most effective at discriminating that tissue
type from the other two types. How the RadViz display achieves this class discrimination
for each clinical sample or point in the display is as follows. The genes are selected from
the large starting population by an algorithm that uses varying statistical criteria to rank
the genes best able to discriminate all the tissues (points) into their known class. The loca-
tion of each tissue sample point shown in the display is determined using a Hooke's
spring paradigm to vectorially sum its measured mRNA expression magnitudes from the
microarray biosensor experiment for all 15 genes arrayed at their specific positions around
the circle. The data points for all members of the three tissue type samples from the larger
data set we analyzed are shown in the RadViz display. By and large, for the result shown
q
FIGURE 1.52
Supervised learning using RadViz for the three-class classification of cancer tissue biopsy samples into normal
(squares), adenocarcinoma (circles), and squamous cell carcinoma (triangles) tissue classes. The RadViz algorithm
classified the known cancer cell tissue types based upon pathologists's clinical determinations using microarray
mRNA transcript level determinations on each tissue sample to carry out the classification by selecting those genes
whose mRNA transcripts are best at classification. The genes whose mRNA transcript levels were most effective
and met the statistical criterion for classification are shown in each tissue class sector. In the RadViz algorithm, the
positions of these genes are what pulls each of the tissues (points) into their respective positions, forming a visual
output of the supervised clustering algorithm. Reprinted from McCarthy, J.F., Marx, K.A., Hoffman, P.E., Gee,
A.G., O'Neil, P., Ujwal, M.L., Hotchkiss, J. (2004). Applications of Machine Learning and High-Dimensional
Visualization in Cancer Detection, Diagnosis and Mangement. In: Umar, A., Kapetanovic, I., Khan, J., eds.
The
Applications of Bioinformatics in Cancer Detection, Ann. N.Y. Acad. Sci.
1020:239-262.
