Segmentation of cardiac anatomical structures in cardiac magnetic resonance\nimages (CMRI) is a prerequisite for automatic diagnosis and prognosis of\ncardiovascular diseases. To increase robustness and performance of segmentation\nmethods this study combines automatic segmentation and assessment of\nsegmentation uncertainty in CMRI to detect image regions containing local\nsegmentation failures. Three state-of-the-art convolutional neural networks\n(CNN) were trained to automatically segment cardiac anatomical structures and\nobtain two measures of predictive uncertainty: entropy and a measure derived by\nMC-dropout. Thereafter, using the uncertainties another CNN was trained to\ndetect local segmentation failures that potentially need correction by an\nexpert. Finally, manual correction of the detected regions was simulated. Using\npublicly available CMR scans from the MICCAI 2017 ACDC challenge, the impact of\nCNN architecture and loss function for segmentation, and the uncertainty\nmeasure was investigated. Performance was evaluated using the Dice coefficient\nand 3D Hausdorff distance between manual and automatic segmentation. The\nexperiments reveal that combining automatic segmentation with simulated manual\ncorrection of detected segmentation failures leads to statistically significant\nperformance increase.\n