Echo-SyncNet

Echo-SyncNet synchronizes cardiac views in echocardiography without electrocardiogram (ECG) inputs by using self-supervised learning to temporally align cross-sectional 2D echo cine series for cardiac-phase correspondence.

Key Features:

• Self-supervised learning framework: Learns temporal alignment directly from input cine data using supervisory signals derived from the videos rather than ECG.
• Intra-view self-supervision: Utilizes temporal intra-view signals based on frame ordering and spatial intra-view signals based on spatial similarities between adjacent frames within a single cine.
• Inter-view self-supervision: Promotes similar embeddings for frames captured at identical cardiac phases across different echo views.
• Encoder-style convolutional neural network (CNN): Processes cardiac ultrasound videos to produce low-dimensional, feature-rich embedding sequences used for synchronization.
• Generalization across cardiac views: Operates without explicit assumptions about specific cardiac view types, enabling generalization to unseen views.

Scientific Applications:

• Synchronization of perpendicular cardiac views: Synchronized Apical 2 chamber and Apical 4 chamber views using data from 998 patients.
• Comparison to supervised methods: Learned representations outperformed a supervised deep learning method optimized for detecting fine-grained cardiac cycle phases in a study of 3,070 patients.
• One-shot learning for key-frame detection: Achieved key-frame detection by synchronizing validation patient studies with a single labeled reference cine across 1,188 patient studies without model fine-tuning.

Methodology:

Self-supervised training using intra-view temporal and spatial signals and inter-view embedding alignment, implemented with an encoder-style CNN that produces low-dimensional embeddings from 2D echocardiography cine sequences.