Breast Cancer Detection Using Transfer Learning in Convolutional Neural Networks Open Access
Computer-aided diagnosis (CAD) in mammography can improve treatment outcomes for breast cancer and provide greater survival times for patients. For breast cancer detection, the Convolutional Neural Network (CNN) can extract features from mammographic images automatically and then do classification. To train the CNN from scratch, however, requires a large number of labeled images. Such a requirement often is infeasible for some kinds of medical image data such as mammographic tumor images. A promising solution is to extract features by reusing a pretrained CNN model that has been trained by very large image datasets from other fields; alternatively, we could re-train (fine-tune) such a model using a limited number of labeled medical images. This approach is also called transfer learning. In this study, we applied the pre-trained VGG-16 model to extract features from input mammographic images and used those features to train a Neural Network (NN)-classifier. We firstly downloaded mammographic images from the DDSM database and cropped the Regions of Interest (ROIs) of given abnormal areas as ground-truth information. We used the ROIs instead of the entire images to train neural networks. The structure of CNN in transfer learning was the combination of the 13 convolutional layers in pre-trained VGG16 model with a simple full-connected (FC) layer. The weights in the FC layer were randomly initialized and updated by training; other weights were not changed. We used 1300 abnormal ROIs and 1300 normal ROIs. All ROIs were randomly selected and shuffled in class sets. After 100 epochs, the average of 10-fold cross validation accuracy converged at about 0.905 for abnormal vs. normal classifications on mammograms, with no obvious overfitting. Our best model could reach 0.950 accuracy for the abnormal vs. normal case, and the area under the receiver operating characteristic curve was 0.96. This study shows that applying transfer learning in CNN can detect breast cancer from a mammogram, and that training a NN-classifier by feature extraction is a feasible method in transfer learning. Our research is important because it provides a novel technique to improve mammographic detection. Compared with other studies in this field, this study used a different pre-trained model, simpler classification architecture, and classifier, and used more images (2600), and performed at least as well.
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.