EchoNet-Dynamic

Cardiac function estimation from echocardiographic data generally involves supervised learning that is costly in terms of clinical labels. In this work, we present a novel self-supervised learning scheme based on Masked Volume Modelling (MVM), motivated by Masked Autoencoders and SimMIM. The proposed framework aims to learn latent representations from volume time series obtained from echocardiograms. In contrast to earlier works, which treat raw 2D videos or static images, we treat cardiac volume data as 1D signals, masking parts of the time series and then reconstructing them in a Transformer encoder-decoder framework. This technique eliminates the need for labeled data, enabling strong downstream efficacy on ejection fraction (EF) prediction, performing unsupervised clustering, and stratifying illnesses. Our work is two-fold: (1) We establish the mathematical underpinnings of MVM through theoretical errorbounds on reconstruction and convergence guarantees, and (2) We set up a comparative platform for MVM and traditional signal processing methods\u2014like Fourier and Wavelet transforms\u2014 to exhibit the specialties of learned representations for cardiac signal reconstruction. Our model performs better thanconventional CNNs and LSTMs and offers both physiological interpretability and computational efficiency. In addition, we provide a comparison with state-of-the-art approaches and highlight our contributions, including phase-aware masking and interpretability by SHAP analysis.