Cardiac landmark localization on axial stacks without dedicated ground truth

Motivation: In cardiac magnetic resonance imaging, accurately identifying anatomical landmarks is crucial for correctly prescribing the standard views needed to navigate the anatomy. Goal(s): This study introduces a novel deep learning training approach for cardiac magnetic resonance imaging that reduces dependence on manual annotations. Approach: By leveraging outputs from pre-trained long-axis models as surrogate ground truth, the method simplifies database creation and maintenance for training networks. Results: Tested on 578 exams and validated on 100 clinical cases, the model achieved comparable accuracy to manually trained models, with minor deviations mainly due to complex pathologies. Impact: The proposed approach eliminates the need for manual annotations for the training of some cardiac models. Using outputs from existing long-axis models as surrogate ground truth simplifies the creation and maintenance of the training database.