Automatic Whole-Heart Segmentation in Congenital Heart Disease Using Deeply-Supervised 3D FCN

Accurate whole-heart segmentation plays an important role in the surgical planning for heart defects such as congenital heart disease (CHD). In this work, we propose a deep learning method for automatic whole-heart segmentation in cardiac magnetic resonance (CMR) images with CHD. First, we start with a 3D fully convolutional network (3D FCN) in order to ensure an efficient voxel-wise labeling. Then we introduce dilated convolutional layers (3D-HOL layers) into the base-line model to expand its receptive field, so as to make better use of the spatial information. Last, we employ deeply-supervised pathways to accelerate training and exploit multi-scale information. We evaluate the proposed method on 3D CMR images from the dataset of the HVSMR 2016 Challenge. The results of controlled experiments demonstrate the efficacy of the proposed 3D-HOL layers and deeply-supervised pathways. We achieve an average Dice score of 80.1% in training (5-fold cross-validation) and 69.5% in testing.