Deep Learning Models for Detecting Breast Cancer : A Comparative Study

Abstract: Breast cancer is the second leading cause of death for women worldwide. As early detection is vital, new methods for facilitating accurate diagnosis are continuously being developed. This project investigates Deep Learning algorithms for detecting breast cancer, specifically two Convolutional Neural Networks (CNNs), comparing their accuracy in correctly classifying lesions in mammograms as benign or malignant. The models used in this project were Breast Cancer Classifier (BCC) and the Globally-aware Multiple Instance Classifier (GMIC). The degree of which the accuracy differs for different levels of breast density, a measurement describing the nature of the tissue of the breast, was also studied. In order to compare the models, they were tested on the same dataset of mammograms. Performance was evaluated by visualising data, and using statistic methods such as confusion matrices and clustering. Both models achieved high accuracy (accuracy BCC = 0.875 and accuracy GMIC = 0.903), but some questions were raised by the results, as higher benign predictions were over-represented. Some possible reasons for the over-predicting of benignancy were discussed, such as class imbalance in the testing dataset. As GMIC did not predict any malignancy, the implications of both false positive and false negative diagnoses were considered when evaluating the models. This led to discarding the GMIC as the best model for correctly identifying lesions in mammograms, despite having the highest accuracy score. The analysis of the results from a breast density perspective did not reveal any significant correlation between breast density and accuracy. Future research could, for example, explore transfer learning to allow for less specific image pre-processing methodology, and curating well-balanced datasets for both training and testing purposes.