Investigating the effect of class-imbalance on convolutional neural networks for angiodysplasia detection

Abstract

In this thesis, we look at a deep learning approach to AD detection and focus specifically on the problem of class imbalance, which arises from the fact that lesions only occupy a small part of the images, by analyzing how weighting of the loss function can help address this issue. Balancing the weights of the foreground and background class in the cost-function was found to be crucial to achieve good segmentation results.