Loss, post-processing and standard architecture improvements of liver deep learning segmentation from Computed Tomography and magnetic resonance

Abstract

As deep learning is increasingly applied to segmentation of organs from Computed Tomography (CT) and Magnetic Resonance Imaging (MRI) sequences, we should understand the importance of certain operations that can improve the quality of results. For segmentation of the liver from those sequences, we quantify the improvement achieved with segmentation network, loss function and post-processing steps. Our results on a publicly available dataset show an improvement of 11% points (pp) by using DeepLabV3 instead of UNet or FCN, 4 pp by applying post-processing operations and 2pp using the top-performing loss function. The conclusions of this work help researchers and practitioners choosing the network and loss function and implementing effective post-processing operations.