The problem of cancer detection and diagnosis is ideally suited for machine learning methods. The goal is to develop a diagnostic model of biological measures extracted from imaging methods, such as mammography, or pathology assays, such as immunohistochemistry. Good predictive models could then be implemented in an expert or intelligent system that would complement the skills of a trained radiologist or pathologist, for example.
This paper proposes an artificial neural network (ANN) framework for diagnosing breast cancer in presymptomatic women using mammography data. The author used a three-layer ANN with back-propagation and applied it to the mammography data for more than 800 women who had mammograms and were subsequently diagnosed with breast cancer following biopsy. Inputs or features modeled by the ANN included the age of the subject and the shape, margin, and density of the mass detected by the mammogram. Also included was the breast imaging-reporting and data system (BI-RADS) score that radiologists use to classify mammograms by severity. The class being predicted was whether each confirmed breast cancer was benign or malignant. A form of cross-validation using holdout data was used to assess performance. Sensitivity, specificity, and receiver operating characteristic (ROC) curves were used to assess ANN performance. The author shows that the ANN predicted well with an ROC area under the curve (AUC) of more than 0.87. Age and BI-RADS score seemed to be the best predictors.
The study demonstrates how an ANN could be used in the context of cancer diagnosis. This was a relatively simple study that could be expanded in numerous different ways. First, it would be nice to evaluate the statistical significance of the ANN classifier. This could be accomplished by estimating a 95 percent confidence interval for the ROC curve and comparing that to the baseline curve expected by chance. Further, permutation testing could be used to assess the likelihood of observing an ROC AUC of 0.87, assuming the null hypothesis of no association is true. These simple methods would add some statistical rigor to the analysis and provide some additional information that would be helpful to physicians who might be interested in adopting a similar approach.
Second, it would be nice to know what effect changing the ANN architecture or other parameters would have on the analysis. The performance of multiple different ANN approaches could be compared in a cross-validation framework using a statistical test such as the resampled t-test. Third, it would be interesting to add additional features to the analysis. For example, additional factors such as body mass index or smoking status could be informative for diagnosing cancer along with radiological features from mammograms. It might also be productive to add measured genetic factors. Indeed, whole genome sequencing is upon us, and we will soon be able to have all the genomic information for any subject at little cost.
Finally, there are steps that could be taken to provide a statistical interpretation of the ANN model. For example, what are the relationships between the nodes in each of the hidden layers? Are they independent, redundant, or synergistic? Measures of entropy could be used to assess these relationships. This would provide a much clearer picture of where the information in this ANN is coming from. Additional analyses of the weights in the ANN might also be informative.
Cancer diagnosis is a rich area of investigation, and there is clearly a central role for machine learning, pattern recognition, and related areas in data mining.