Multi-class semantic segmentation and volume calculation of cardiac CT scans using convolutional neural networks

Abstract

Computed tomography (CT) scans from the MICCAI 2017 Multi-Modality Whole Heart Segmentation Challenge dataset were segmented into seven anatomical substructures (left and right ventricles, left and right atria, myocardium, aorta, and pulmonary artery) using the SpatialConfiguration-Net (SCN). The SCN utilizes a U-Net convolutional network architecture that is designed for biomedical image segmentation. The performance of the neural network was evaluated by cross validation on the training data by the Dice Similarity Coefficient over all heart substructures. Improved segmentation of most of the substructures was observed over other proposed convolutional neural network pipelines. While the estimated volumes of the segmented substructures are consistent with the manual annotations, the SCN performed sub-optimally in detecting gaps between heart substructures due to the loss of resolution caused by multilayer down-sampling and up-sampling of the images.