Improved Endoscopic Polyp Classification using GAN Generated Synthetic Data Augmentation

Early diagnosis of cancer in polyps detected in the endoscopic video frames helps in better prognosis and clinical management. For this, the polyp regions are exhaustively analyzed by an endoscopist. In this paper, an automated polyp classifier in a deep learning framework is proposed. As the availability of ground truth data for colonic polyp is always in paucity, data augmentation is indispensable in such task. Our work proposes the use of Generative adversial networks (GANs) for synthetic data generation. For classification, a CNN is trained which discriminate between normal (benign) and cancer (malignanat) polyps. Experiments carried on two databases prove that the proposed data augmentation technique can efficiently be used in the classification of colonic polyps. Also, our proposed method compares the performance achieved using classical augmentation approach which is generally considered in limited data scenario. Experimental results show that the classification accuracy is competitive to the state-of-the-art methods.