Fig. 1. Workflow of the two-stage paradigm for automatic screening and diagnostics of CVDs.

For each patient, the screening model takes cine movies as inputs and outputs the binary classification to detect cardiac anomaly. Second, the patient suspected of cardiac anomaly undergoes LGE imaging, while the diagnostic model integrates both cine and LGE to output their CVD class. The AI models comprised four VST blocks to analyze the CMR sequences using the 3D-shifted window self-attention (WSA) mechanism. MLP, multilayer perceptron; norm, normalization; num, number; layer l and layer l + 1 indicate two consecutive layers.