Feed

The automatic estimation of left ventricular ejection fraction (LVEF) from echocardiograms is crucial for the diagnosis of cardiovascular diseases. However, constrained by the inherent speckle noise of ultrasound imaging, the complex temporal dynamics of the cardiac cycle, and the continuous distribution characteristics of physiological indicators, current methods often struggle to extract feature representations that exhibit both robustness and semantic continuity. To address this challenge, we propose a unified structural contrastive learning framework for LVEF applications, designed to seamlessly integrate two complementary structural constraints into the feature learning process through a unifying mathematical formulation. The framework dynamically redefines the sampling strategy for positive and negative pairs. Specifically, it incorporates (1) a temporal consistency constraint via key region masking and (2) reconstruction-based denoising during the self-supervised pre-training phase to capture noise-resistant and stable anatomical structures. During the fine-tuning phase, we further introduce an EF-aware continuous constraint to guide the feature space along the LVEF gradient to form an ordered continuous manifold. Extensive experiments on the EchoNet-Dynamic dataset demonstrate that with only 36 input frames, our method achieves a mean absolute error (MAE) of 3.84, a root mean square error (RMSE) of 5.12, and an R² of 0.82. Furthermore, evaluating its clinical utility using an automated medical report generation pipeline shows that our module improves downstream diagnosis accuracy to 88.21% and enhances report reliability, particularly near critical diagnostic thresholds. Without relying on large-scale external data for pre-training, the proposed method achieves highly competitive performance compared to current state-of-the-art approaches, providing a promising tool for precise cardiac function assessment in clinical settings.