Mammography Data Augmentation Using ACGAN
Yijiang Fan, Jiajia Jiao · 2021
As high morbidity and mortality of breast cancer, early diagnosis plays a key role in improving survival rate. So it is significant to develop breast cancer detection techniques. Nowadays, Convolutional Neural Networks (CNN) have been proven to detect breast cancer well, but face two challenges: lack of large labelled datasets and unbalanced distribution of incident categories. Therefore, we propose BreastGAN for further data augmentation via using an Auxiliary Classifier Generative Adversarial Network (ACGAN), in order to generate labeled images. Integrating self-attention and spectral normalization components in the generator and discriminator respectively, BreastGAN can generate higher resolution images and more accurate labels. Additionally, label smoothing can further improve the accuracy of labels by increasing the distance between different labels and reduce the distance within same labels. In our experiments, BreastGAN is also evaluated by the accuracy of the trained classification model with new synthesized datasets, not limited to the generated images quality themselves. The results show that our proposed model achieves up to higher IS (56.54) and lower FID (21.25) of image quality, 4.9% higher accuracy of classification model over latest ACGAN.