GAN based Data Augmentation for Enhanced Tumor Classification

With the incredible breakthrough of medical imaging intertwined with the computer aided diagnosis and artificial intelligence paved a way for the early detection of tumor. Though Deep Neural Networks is the new paradigm in the field of computer vision yet, they are highly reliable on large dataset to avoid overfitting. To overcome this problem, our work focuses on data augmentation, a quintessential solution to handle the inadequate medical data. This paper involves conventional data augmentation using affine transformations. The conventionally augmented data are further synthesized using General Adversarial Networks (GANs). These methods are employed on the benchmark breast tumor datasets namely MIAS, DDSM and INBreast. The classification of benchmark dataset resulted in an accuracy of 69.85%, which are increased by the conventional data augmentation techniques to 88%. The synthesized image when merged with the original benchmark dataset and the conventional augmented images outperformed the others with an increased accuracy of 94%.