Hybrid-Frequency Feature Evolution Network for Endoscopic Ultrasound Image Segmentation

The segmentation of lesions in endoscopic ultrasound images is essential for clinical diagnosis. However, such as the low contrast in ultrasound images, patient-specific factors, and inter-case variations, make accurate segmentation highly challenging. To address these issues, we propose a novel Hybrid-frequency Feature Evolution Network (HFENet) for EUS image segmentation. Specifically, we propose a Hybrid-frequency Feature Evolution and Supervision (HFES) module that ensures consistent segmentation across multiple scales, mitigating the impact of patient-specific factors and inter-case variability. Additionally, by integrating wavelet transform, multi-scale features are decomposed into low- and high-frequency components to capture global structures and fine details, respectively. A Selective Co-Evolution (SCE) module is also proposed to adaptively balance these components to enhance feature representation. Experimental results on three EUS datasets show that our HFENet outperforms state-of-the-art segmentation methods.