The machine learning literature in breast cancer research is vast (Cruz and Wishart, 2006; Kourou et al., 2015). Here we summarize one study that highlights the potential for machine learning to aid in breast cancer diagnosis, when used in conjunction with mammography and EMRs. The study is from the University of Wisconsin, Madison and consisted of 48,744 mammograms (Ayer et al., 2010).

For each mammogram, information about 36 categorical variables was collected, including clinical data (age, past medical history, family history) and mammographic findings, such as tumor mass characteristics, surrounding skin and nipple characteristics, lymph node examination, and calcification characteristics. An artificial neural network consisting of 1,000 hidden layers was trained and the generated labels were compared to the true label of benign versus malignant. Eight radiologists also reviewed varying numbers of mammograms and classified the scan as benign versus malignant. The total AUC of the neural network was 0.965, while that of the radiologists was 0.939. This study, combined with other studies we will read about in the final chapter, demonstrate how ML can be used as an effective tool in the fight against cancer: