Automated Scar Segmentation From Cardiac Magnetic Resonance-Late Gadolinium Enhancement Images Using a Deep-Learning Approach

Sara Moccia1, Riccardo Banali2, Chiara Martini3, Giuseppe Muscogiuri4, Gianluca Pontone4, Mauro Pepi4, Enrico Caiani5
1Department of Electronics, Information and Bioengineering, Politecnico di Milano / Department of Advanced Robotics, Istituto Italiano di Tecnologia, 2Department of electronics, information and bioengineering Politecnico di Milano, 3Department of Medicine and Surgery University of Parma, 4Centro Cardiologico Monzino IRCCS, 5Politecnico di Milano


Aim. Location and extension of myocardial scar following ischemic-heart disease are strong predictors of ventricular remodeling, cardiac dysfunction and mortality. Our aim was to assess quantitatively the presence of myocardial scar from cardiac-magnetic-resonance (CMR) with late-Gadolinium-enhancement (LGE) images using a deep-learning (DL) approach. Methods. 250 CMR-LGE images of 30 patients (26 men and 4 women) with ischemic-heart disease were analyzed retrospectively. Images were acquired at Centro Cardiologico Monzino in Milan, Italy. All patients were imaged with a 1.5-T scanner (Discovery MR450, GE Healthcare, Milwaukee, WI). Scar tissue was present in all patients, but only in 215 CMR-LGE images. Scar segmentation was performed automatically with a DL approach based on ENet, a deep fully-convolutional neural network (FCNN). We investigated three different ENet configurations. The first configuration (C1) exploited ENet to retrieve directly scar segmentation from the CMR-LGE images. The second (C2) and third (C3) configurations performed scar segmentation in the myocardial region, which was previously obtained in a manual or automatic way with a state-of-the-art DL method. Each configuration was evaluated using leave-one-patient-out cross validation. Data augmentation was performed. For FCNN training and evaluation, gold-standard scar contour and myocardial area were obtained with manual segmentation performed by one expert clinician. Results. With the best-performing configuration (C2), scar-segmentation median accuracy (Acc) and Dice similarity coefficient (DSC) were 97% (inter-quartile (IQR) range = 4%) and 71% (IQR = 32%), respectively. C2 outperformed both C1 (Acc = 96%, DSC = 55%) and C3 (Acc = 95%, DSC = 51%), where Acc and DSC for myocardial segmentation were 94% and 86%, respectively.
Conclusions. DL approaches using ENet are promising in automatically segmenting scars in CMR-LGE images, achieving higher performance when limiting the search area to the manually-defined myocardial region. A complete automated approach still suffers from cumulative errors from both myocardial and scar segmentation.