Deep Learning and Image Pyramids for Aorta Segmentation

Abstract

Deep learning techniques have been employed in a variety of use cases for computer vision tasks in general, and medical image segmentation in particular, with U-Net being a notable example. In this thesis, and through the use-case of aorta segmentation in full body CT scans, we investigate the combination of image pyramids and U-Net as a deep-learning architecture for medical image segmentation. Our experiments show that Gaussian and Laplacian pyramids provide equal performance when used as output of a U-Net model, we also show that extensions of such model provide only small benefits in restricted conditions. Finally, we propose an image-pyramid-based model that provides comparable results to the state-ofthe-art nnU-Net, with a difference of only 0.3% in Dice score, while having 120× less parameters, and requiring 20× less time to train