3D U-Net based method for fast segmentation of whole heart from CT images

Abstract

There is an increasing number of clinical applications where deep learning plays an important role. Heart chamber segmentation enables delineation of anatomical structures of heart and it is a prerequisite for a wide range of clinical applications. 3D U-Net architecture consistently achieves the highest scores in various medical imaging challenges. NoNewNet architecture is a modification of the 3D U-Net architecture which was shown to outperform the original 3D U-Net and was recently implemented inside the NiftyNet package. In this paper we demonstrate that with the properly trained NoNewNet network and NiftyNet we can outperform the current state-of-the-art networks. The evaluation of the trained network was performed on 20 3D CT images from MICCAI 2017 Multi-Modality Whole Heart Segmentation Challenge dataset using five-fold cross-validation. We experimentally prove that border size can significantly reduce inference time without affecting segmentation accuracy. Additionally, we provide the discussion of the effects of some of the NiftyNet configuration parameters on the performance of the network.