Domain Shift in Echocardiography: Interpretable Quantification and Prediction of Cross-Dataset Left Ventricular Segmentation
The research investigated how echocardiographic left ventricular segmentation models generalize across different datasets by analyzing three types of data representations: handcrafted features, VAE latent features, and segmentation-derived features. It found that geometry-aware preprocessing helps mitigate domain shift caused by framing inconsistencies rather than just acoustic differences. The study provides representation-specific measures to predict transfer degradation before clinical deployment.