In many practical applications, datasets used for training neural networks
are highly unbalanced or scarce in volume. Examples can be
found in medical imaging, where data from common diseases are significanlty more than that of rare diseases, or in financial transactions with much more healthy transactions than fradulent ones, or in the
analysis of network traces with hackers traffic constituting only a small amount of communications. Training deep neural networks with unbalanced data tends to bias the network towards the classes that contain more data.
In our research, we address these shortcomings by data augmentation:
Data augmentation via radial transformation
Salehinejad H, Valaee S, Dowdell T, Barfett J., "Image Augmentation using Radial Transform for Training Deep Neural Networks", IEEE International Conference on Acoustics, Speech and Signal Processing (IEEE ICASSP), 2018. Link to Arxiv
Salehinejad H, Naqvi S, Colak E, Barfett J, Valaee S., "Unsupervised Semantic Segmentation of Kidneys Using Radial Transform Sampling on Limited Images," IEEE Global Conference on Signal and Information Processing (IEEE GlobalSIP), California, USA, 2018.
Salehinejad H, Naqvi S, Colak E, Barfett J, Valaee S. "Cylindrical Transform: 3D Semantic Segmentation of Kidneys with Limited Annotated Images," IEEE Global Conference on Signal and Information Processing (IEEE GlobalSIP), California, USA, 2018.
Radially Transformed Images
Data Augmentation via Generative Adversarial Networks
Salehinejad H, Valaee S, Dowdell T, Colak E, Barfett J. "Generalization of Deep Neural Networks for Chest Pathology Classification in X-Rays Using Generative Adversarial Networks", IEEE International Conference on Acoustics, Speech and Signal Processing (IEEE ICASSP), 2018.
Link to Arxiv
Salehinejad H, Dowdell T, Colak E, Barfett J, Valaee S., "Synthesizing Chest Radiograph Pathology for Training Deep Convolutional Neural Networks", Accepted in IEEE Transactions on Medical Imaging, 2018.
