"Monotonicity and Logical Analysis of Data: A Mechanism for the Evaluation of Mammographic Clinical data"
Proceedings of the 13-th Symposium for Computer Applications in Radiology (SCAR),
Denver, CO, June 6-9, 1996, pp. 191-196.
by Kovalerchuk, B., E. Triantaphyllou, and J.F. Ruiz
Abstract:
Computer assisted diagnosis has become one of the promising methods for improving the accuracy
and early detection of breast cancer. Standard back-propagation neural networks are presently
very popular diagnostic tools, but this approach does not inform the physician user
on how a conclusion was reached. Some promising results have obtained with rule based
techniques. This method has the advantage of providing the physician with a
tool that promotes consistency and accuracy. However, these (and most other) models suffer
from relatively small training sample sets which in term limit statistical significance.
An approach called logical analysis of data (LAD), and which is based on
inferring discriminant Boolean functions, has the potential to overcome these weaknesses. By
discovering logical relationships in existing classes of disjoint observations, the method
can improve the understanding of the diagnostic process The statistical significance problem
is eliminated by exploiting the property of monotonicity that exists within mammographic evaluation and
interpretation.
The resultant discriminant functions could be used many ways.
In the evaluation of a new "problem case" the radiologist could use
these functions for that case to draw a diagnostic conclusion. Alternatively, with a set
of "gold standard" test cases, the functions
could be used as a reproducible testing mechanism. Each radiologist could determine his/her
own function, compare it to the gold standard and thereby identify areas of strength or weakness.
Progress or improvement over time could be objectively measured.
Key Words:
Digital mammography, breasr cancer, data mining, knowledge discovery.
