Comparison of regularization techniques for DCNN-based abdominal aortic aneurysm segmentation

This study compares several state-of-the-art regularization methods applicable to aortic aneurysm segmentation likelihood maps provided by a Deep Convolutional Neural Network (DCNN). These algorithms vary from simple Otsu's thresholding and K-Means clustering, to more complex Level-sets and Conditional Random Fields. Experiments demonstrate that K-means yields the best results for the current application, which poses the question about the need to employ a more sophisticated approach for post-processing the output probability maps.