Frequency-Aware Spatiotemporal Modeling with Cross-Frequency Attention for Echocardiography Left Ventricle Segmentation

Accurate segmentation of the left ventricle (LV) in echocardiography videos is essential for cardiac function assessment but remains challenging due to complex motion patterns, temporal dependencies, and noisy artifacts. In this work, we propose a novel frequency-aware spatiotemporal modeling network that captures both video-level and frame-level frequency characteristics for robust LV segmentation. At the video level, the extracted spatiotemporal feature tensor is decomposed into low-frequency components that capture periodic cardiac dynamics and high-frequency components that retain fine structural details. A cross-frequency attention mechanism enables high-frequency queries and values to interact with low-frequency keys, enhancing boundary learning under the guidance of global motion consistency. At the frame level, we utilize joint spatial-frequency feature extraction in the encoder and adaptively fuse high- and low-frequency components through skip connections in the decoder, enhancing segmentation precision and structural fidelity for each frame. Experiments on the CAMUS dataset demonstrate that our method outperforms state-of-the-art approaches, effectively leveraging both temporal coherence and frame-level detail for accurate LV segmentation.