Cascaded Volumetric Fully Convolutional Networks for Whole-Heart and Great Vessel 3D segmentation

Abstract

The absence of cardiac chambers, holes in the heart and abnormal connections cause the death of hundreds of people every year. The clinical team involved in the diagnosis and treatment decisions for congenital heart disease (CHD) must be consistent in their choices. Therefore, in cases of CHD there is a need for a system that is capable of segmenting the whole heart and large blood vessels in 3D efficiently, quickly, and accurately. This article proposes to fill this need by using Cascaded Volumetric Fully Convolutional Networks. The approach proposes the use of two Volumetric Fully Convolution Networks (V-Net) in sequence. The first network aims at locating the cardiac region, while the second segment the substructures of the cardiac area and the great vessels. Both networks are trained with the 2016 data set from the MICCAI Workshop on Whole-Heart and Great Vessel Segmentation of 3D Cardiovascular MRI in Congenital Heart Disease (HVSMR). The experimental results show that the proposed method has a promising potential in decision-making in CHD cases (from MR images). The approach obtained on average 98.15% for Accuracy, 94.89% for Precision, 98.81% for Specificity Coefficient, 94.27% for Sensitivity Coefficient, 93.24% for Matthews Coefficient, 80.65% for Jaccard Index, 94.20% for Dice Coefficient, and 1.61 for the Hausdorff Distance. The proposed method enables the visualization and iteration of the segmented volume in 3D so that the doctor can analyze the entire structure of the heart along with the circulatory network.