We propose a two-stage weighted ensemble deep-learning method to predict cardiac left-ventricular edema and scar labels from multi-modal MR images (LGE, T2-weighted, and bSSFP) provided by the MyoPS++ challenge hosted by MICCAI 2024. In the first stage, a 2D nnU-Net is trained on all three modalities to predict the myocardium mask, which defines a bounding box and serves as input to the networks in the second stage. In the second stage, three networks are used to predict healthy myocardium, scar, and edema labels from the multi-modal images, and the final labels are obtained from the weighted average of the three individual predictions. Specifically, a 2D nnU-Net is trained on two-channel inputs, LGE and binary myocardium (N = 1957), from six available centers. A 2D nnU-Net is trained on 4-channel inputs from all modalities available from two centers (N = 379). Lastly, a 3D nnU-Net is trained on volumes of all modalities from two centers (N = 85). The weights were optimized on the validation dataset (N = 50), achieving challenge validation Dice Scores of 0.599 and 0.660 for scar and scar + edema, respectively.