SemanticST: Semantics-enhanced Spatio-Temporal Modeling for Ejection Fraction Estimation in Echocardiography

Estimating left ventricular Ejection Fraction (EF) from echocardiography is critical for cardiac systolic function assessment and clinical risk stratification. Unfortunately, existing EF estimation methods are limited by (i) insufficient modeling of temporal clues embedded in video frames and (ii) inadequate exploitation of clinical semantics in easily accessible textual reports. In this study, we propose a unified segmentation and EF estimation framework with Semantics-enhanced spatio-temporal modeling (named SemanticST), integrating spatio-temporal consistency modeling with structured clinical semantic priors. Our SemanticST introduces a spatio-temporal & text-guided neighborhood correlation mining (STT-NCM) encoder, which captures both short- and long-range temporal dependencies via text-modulated 3D neighborhood attention. Further, a text-guided pixel-level semantic projection module (TextSP) is designed to map the key clinical cues extracted by a large language model (LLM) into pixel-level guidance features, enabling the alignment of semantic priors with visual context for optimized EF estimation. Extensive experiments on two public datasets (CAMUS and EchoNet-Dynamic) demonstrate that our SemanticST outperforms state-of-the-art methods in segmentation accuracy, temporal consistency, and EF estimation correlation.