The core of the project is the University of Wisconsin Original Breast Cancer Data Set. An optimized version of that dataset was created to improve the performance of the cloud service during initialization. The contents of the data did not change; however, the way it was represented did. A text file version of the following document is accessed by the program.
The following table is the summary data for the commercial neural network trials. For each network, ten percent of the data was used to validate the network. As can be seen, two networks experienced consistent success while the third experienced erratic success. While the overall success rates for the successful implementation fall only slightly below the custom network's rates, the sensitivity success of the custom network is unparalleled by the commercial products. On sheet two, a more detailed breakdown of the most successful network's results demonstrates the average of the sensitivity of the commercial network was 4.97% lower than the sensitivity of the custom network.
Above are the raw results for the trial of the custom network. The custom network was tested extensively; over 6,800 tests were run. As can be seen, the network is tested by removing one sample from the training set. Ten different networks diagnose this sample, and if all networks get the same result, a conclusive diagnosis is returned. Otherwise the mass is declared inconclusive. When the network was finished testing, overall efficiency was 97.41%. Malignant sensitivity was 99.11%.
On the second tab of the spreadsheet are the results for trials run when all samples were included in the training set. When the network predictive abilities were assessed, 100% of diagnosed masses were correctly detected and only four masses were ruled inconclusive.
Over 7.6 million tests were conducted on the network to determine the effect sample size had on predictive success. The network was found to experience more predictive success when the training bank was largest. As more samples were included in training, not only did the network diagnose more samples correctly, but it also felt comfortable identifying some border-line cases, so less samples were declared inconclusive.
In the data table above, each of the networks was run through 1,500 iterations. The networks are ordered in decreasing size of training set and thus, increasing size of validating set.