Mammogram-Based Cancer Detection Using Deep Convolutional Neural Networks

In recent years, applying deep learning to medical images has experienced a surge but often comes with limitations related to the datasets: publicly available datasets have the drawback of being relatively small compared to other datasets used in image recognition tasks. We show multiple findings in our work: the immense power of Deep Convolutional Neural Networks even when applied on a small dataset such as the INbreast dataset. We also demonstrate that accuracy is not the only evaluation metric for network performance evaluation: the recall metric should be maximized. We also show the importance of using cross-validation to assure the absence of overfitting during the learning process. Results show an average classification accuracy for 5-fold cross-validation of 80.10% and an average AUC of 0.78. A graphical user interface was implemented in order to be tested by certified radiologists.