Medical usage demands neural networks achieve accuracy with
their diagnosis and reduce malignant false negatives. Building on data
collected by the University of Wisconsin in the early 1990s, this project first
evaluates three modern commercial neural network implementations. Information
regarding potential indicators of breast cancer is quantified in the dataset;
specifically, clump thickness, single epithelial cell size, bare nuclei,
mitoses, and five other attributes. Each network accepts this input to optimize
its hidden nodes and is tested with ten trials. With each trial, a randomly
selected 10% of the dataset is used to assess the predictive power of the
constructed neural network. These commercially created neural networks serve as
a control group.
Development of a custom neural network weights malignant
false negatives and allows for the identification of inconclusive samples (capacities
not available in commercial products.) Additionally, more samples are needed to improve the predictive
capability of the network; therefore, the network has been published in the
cloud, allowing for global submissions and benefit. The cloud service is hosted in the Google App Engine.
The successfully implemented custom network is tested with
6,800 trials. To assure maximum
training, each sample is run through ten trials evaluated by different networks
trained against all other samples.
The custom neural network achieved predictive success of 97.4% with
99.1% sensitivity to malignancy – substantially better than the evaluated
commercial products. Out of the
commercial products, two experienced consistent success while the third
experienced erratic success. The sensitivity to malignancy for the custom
network was 5% higher than the best commercial network’s sensitivity. This
experiment demonstrates modern neural networks can handle outliers and work
with unmodified datasets to identify patterns. In addition, when all data is
used for training, the custom network achieves 100% success with only 4
inconclusive samples, proving the network is more effective with more samples. Additionally, 7.6 million trials were run using different training sample sizes to demonstrate the sensitivity and predictive success improves as the network receives more training samples.
The Global Neural Network Cloud Service for Breast Cancer may be
ready to diagnose actual patients – more global participation is required to
confirm the findings and increase the predictive success on blind samples.