There are many different
types of neural networks, but this project uses back propagation architecture.
The back propagation training technique allows the computer to run through a
situation of specific inputs and assign random weightings of importance to each
input. The network continuously adjusts these weights until the actual outcome
and the projected outcome are the same. In neural networks, there are certain
customizations are common. Among some of these features, learning rate and
learning momentum can be programmed to dictate the network’s speed of learning.
The right amount of hidden nodes must be created so the network receives the
right amount of contributing information.
Breast cancer is a disease
that inflicts one in eight women. In the early 1990s, researchers at the
University of Wisconsin recognized the relevance of improving breast cancer
diagnostics. Under the supervision of Dr. William Wolberg, 699 patients had
fine needle aspirates (FNAs) of their breast masses. FNAs are tests that stick
a needle into a mass to extract cells from the mass for observation. Doctors
then rated different attributes of the cells on a scale of one to ten, one
being indicative of a benign mass and ten being indicative of a malignant
tumor, before they definitively determined the diagnosis of the mass. These
results were published on the UCI Machine Learning Repository for public use.
Due to the sufficient
amount of data, this set has become popular in the realm of ANN application to
medicine. While numerous trials have had success, an interesting point is they use
the diagnostic set of data rather than the original set of data from the
University of Wisconsin. The diagnostic set was stripped of outliers and
doctors weighted the inputs for the neural network instead of using the ANN’s
“brain” to determine input importance. This approach is novel and in some
trials, 100% of malignant tumors were diagnosed correctly.
In 2007, the Universiti
Sains Malaysia and Universiti Malaysia Perlis collaborated to collect data
similar to Wisconsin’s. However, their data related to diagnosing pre-cancerous
stages of breast cancer. This data was utilized to build diagnostic programs.
Seven implementations were tested and results proved encouraging; accuracy
ranged between 75.38% and 100%. The purely AI based programs did not fare as
well as conventionally programmed implementations, but showed promise.
With the success of Malaysia’s programs when
working with raw data, it is time that the Wisconsin set is revisited. Most of
the trials using Wolberg’s data were performed in the 1990s, and with the
improvements to modern technology, successful neural network diagnostics while
working with raw data could be a reality.