Transferring deep neural networks for the differentiation of mammographic breast lesions

Machine learning can help differentiating benign and malignant lesions seen on mammographic images. Conventional models require handcrafting features for lesion representation. Due to insufficient medical instances, the performance of convolutional neural networks (CNNs) can be further increased. This study makes use of transfer learning for mammographic breast lesion diagnosis and deep neural network (DNN) models pre-trained with large-scale natural images are employed. The diagnosis performance is evaluated with the prediction accuracy (ACC) and the area under the curve (AUC) on average. A histologically verified database is analyzed which contains 406 lesions (230 benign and 176 malignant). Involved models include transferred DNNs (GoogLeNet and AlexNet), shallow CNNs (CNN2 and CNN3) that are fully trained with medical instances and boosted by support vector machine (SVM), and two conventional methods which combine handcrafted features and SVM for lesion diagnosis. Experimental results indicate that GoogLeNet achieves the best performance (ACC=0.81, AUC=0.88), followed by AlexNet (ACC=0.79, AUC=0.83) and CNN3 (ACC=0.73, AUC=0.82). Knowledge transfer can improve the mammographic breast cancer diagnosis, while its wide application still requires further verification in medical imaging domain.