Malignant or Benign?Any woman who finds a lump in her breast wants to know the answer to this critical question quickly, accurately, and with minimal invasion.
Can a breast cancer neural network be optimized to improve the success of diagnosis using Fine Needle Aspirates (FNA)? Specifically, can the network be optimized to reduce the number of malignant false negatives while handling original, unformatted data and providing global access via the cloud?
When the Wisconsin data was originally collected in the
early 1990’s, initial neural networks were largely unsuccessful at diagnostics.
Thus, the University republished the data without outliers and with
pre-calculated mathematical weightings for each input. Although this was a
novel approach to gaining success with 90’s technology, current technology has
progressed at a rapid pace and may be able to handle raw, unformatted data.
Artificial neural networks exist in a wide array of types.
Each network is customized to handle its own area of expertise. This project
will explore how different optimization factors affect networks. Four networks
will be tested. The first three networks will be generated by software programs
and will be optimized to their fullest potential. Based on the most efficient
settings and built to incorporate specific features unique to breast mass classification,
a fourth network will be custom-coded in java.
Networks will be extensively tested. Then, the best network
will be tested to see how it functions with different sized training sets to
emulate the effect of more data.