Keywords: Segmentation, Numerical Representation, Image Entropy, Deep-Learning, U-Net, MRI
TL;DR: Lower input image entropy may help with complex segmentation tasks.
Abstract: Deep learning has made major strides in medical imaging segmentation in the last several years for its automated feature extraction. This model fitting process is susceptible to over-fitting, and can benefit from sparsity. Here, we show theoretical and experimental potential of using low-entropy images as sparse input to improve deep learning driven tissue segmentation, using tumor and heart segmentation problems as exemplary cases.