SFMFNet: Spatial‐Frequency Multiscale Fusion Network for Left Ventricle Segmentation in Echocardiography

Echocardiography is a key imaging modality for assessing cardiac structure and function, and accurate left ventricular segmentation is crucial to diagnosing cardiovascular diseases. However, existing left ventricular segmentation methods are limited to feature extraction in a single domain, suffering from inadequate crossdomain fusion and insufficient high/low‐frequency feature interaction. To address this challenge, we propose the Spatial‐Frequency Multiscale Fusion Network (SFMFNet). SFMFNet utilizes multiscale convolution to extract local information, introduces wavelet transform to separate high‐ and low‐frequency features, and adopts Fourier transform to extract global features. The bidirectional crossattention fusion (BCAF) module achieves bidirectional mutual learning, facilitating effective cross–domain interaction. The high‐ and low‐frequency multiscale receptive field fusion (HLMRF) module further integrates these crossdomain features, reducing the receptive field gap and strengthening feature complementarity. SFMFNet accurately segments echocardiographic images and exhibits good generalization. On cardiac acquisitions for multistructure ultrasound segmentation (CAMUS), it achieves Dice scores of 0.9473 and 0.9276 for end‐diastolic (ED) and end‐systolic (ES) phases, and on EchoNet‐Dynamic, it achieves 0.9367 and 0.9160. The results show the proposed network effectively fuses complementary information from spatial and frequency domains, achieving high‐precision segmentation of the target structure in complex ultrasound images, thereby providing reliable support for clinical echocardiography.