Feature space extrapolation for ulcer classification in wireless capsule endoscopy images

Deep convolutional neural network has shown dramatically improved performance not just in computer vision problems but also in various medical imaging tasks. For improved and meaningful result with deep learning approaches, the quality of training dataset is critical. However, in medical imaging applications, collecting full aspects of lesion samples is quite difficult due to the limited number of patients, privacy and right concerns. In this paper, we propose feature space extrapolation for ulcer data augmentation. We build dual encoder network combining two VGG19 nets integrating them in fully connected encoded feature space. Ulcer data is extrapolated in the encoded feature space based on respective closest normal sample. And then, fully connected layers are fine-tuned for final ulcer classification. Experimental evaluation shows our proposed dual encoder network with feature space extrapolation improves ulcer classification.