Cardiac segmentation on magnetic resonance imaging data with deep learning methods

The study compared UNet, ENet, and BoxENet convolutional neural network architectures that provide the various approach of increasing of the receptive field. The analysis employed an Automated Cardiac Diagnosis challenge dataset containing the magnetic resonance imaging data of 150 patients to solve a segmentation problem for left ventricle cavities and myocardium of the right ventricle. We show that while UNet models achieve 5% higher accuracy on the validation dataset than other neural network architectures, ENet and BoxENet can be trained five times faster and require only half the memory than UNet.