*Example synchronization of two perpendicular views, AP4 and AP2.
For each view, we show trajectories of the cines in the embedding space. Using a principal component analysis reduction approach, we reduce the dimensionality of the embedding from 128 to 1 for visualization. We examine the synchronization of an AP4 and AP2 echo cine by comparing three distinct cardiac events: the earliest opening of the mitral valve, the maximum contraction of the left ventricle and the earliest closing of the mitral valve.
*Example synchronization of two other cardiac views, AP4 and PLAX, showing the extensibility and generalizability of the proposed model to other cardiac views.
*Example synchronization of three unique cardiac view angles, AP4, AP2, and PLAX, along with their trajectories in the embedding space.
*Example synchronization of four synched cines captured from AP4, AP2, AP5, and PLAX views along with their trajectories in the embedding space. In this example, the generalizibility of the method even in the views not found in the training set can be observed.
*Example synchronization of cines captured from AP4 and AP2 views along with their trajectories in the embedding space. This experiment shows the robustness of the proposed synchronization method to different frame rates. We first produce downsampled versions of both cines at one-half and one-quarter of their original frame rate. Next, we create embeddings and perform synchronization using various combinations of original, half sampled and quarter sampled pairs.
The need for automatic echo synchronization:
- Calculation of clinical measurements in cardiac echo often require or benefit from having multiple synchronized views or accurately annotated keyframes.
- Traditional methods to synchronize echo rely on external factors such as an electrocardiogram which may not always be available; especially in the point of care setting.
To address these points we propose Echo-SynNet a neural network-based framework for automatic synchronization of cardiac echo. Echo-SynNet is trained using only self-supervised methods and is hence cheap to train or finetune on any dataset.
Echo-SyncNet is trained on a dataset of 3070 unannotated echo studies. We use a multiobjective self-supervised loss, described in detail in our paper, to promote the consistency of embedding features across multiple training samples.