Conclusion

In this experiment, an artificial neural network was implemented to improve on the success of commercial products and reduce the number of malignant false negatives. Through a customized implementation utilizing novel approaches including heavy malignant weighting, an inconclusive diagnosis option, and an artificial neural input layer, success was achieved.  All aspects of the hypothesis were proven correct, specifically:
  1. A custom neural network tuned to heavily weight malignant tendencies improves diagnostic results.
  2. Increasing the number of training samples has a positive correlation to the success rate.  The more samples collected, the more accurate the network will become.
  3. The modern neural network is able to successfully handle outliers and achieve better results than original networks that depended on mathematical formulas identified by research institutions.

With the custom developed artificial network, only two samples out of 681 samples were classified as malignant false negatives. Malignant false negatives are the most dangerous misdiagnosis because they are life-threatening. The custom neural network achieved predictive success of 97.4% with 99.1% sensitivity to malignancy. The custom-crafted neural network was able to improve breast cancer diagnostic results from 94.9% to 99.1% in relation to sensitivity to malignancy versus commercial implementations. This closes 82% of the gap to perfection and brings the utility of the service much closer to viable implementation.



In addition, the accuracy of the network prediction increases and the inconclusive rate decreases as there are more samples. The 99.1% sensitivity occurs when 680 samples are used to train the network and the remaining sample is used to test blindly. When only 340 samples are used for training, predictive success drops to 97.2% and sensitivity drops to 98.4% while inconclusive rates increase to 3.84% from 3.67%. As such, more data allows the network to diagnose some borderline cases and still improve its accuracy. This correlation demonstrates the critical need for global participation and a single repository that can be used by the entire medical community.

Based on the combination of these findings, 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. The impressive results achieved when all data is included in training is a good omen for the potential of the custom neural network. If more data was available to counterbalance outliers, the network could achieve perfection.

This neural network implementation could be leveraged to diagnose other forms of cancer or assist with evaluation of other areas of study as long as information can be clearly captured in numeric terms as inputs.  There are already quality data sets for prostate cancer and ovarian cancer.

Looking forward, the cloud service could be marketed to collect more data globally and improve results. As is shown by the correlation between success and data samples, the network needs more samples to improve accuracy and reduce inconclusive diagnosis. The learning capability proved successful; neural networks can be crafted to prove useful for medical diagnostics.