Neural Multi-Atlas Label Fusion: Application to Cardiac MR Images.

•Our method integrates a feature extraction subnet and a NL-PLF subnet in a single network.•Our method can learn the deep features for computing optimal label fusion weights in a NL-PLF framework.•Our method naturally defines a similarity measure for atlas selection in the testing process.•Our method is evaluated on two cardiac MR datasets for LV segmentation and achieves promising results in various metrics.•Our method was tested on CAP test set of MICCAI 2013 SATA Challenge and ranked highest at the time of writing.