FSA-Net: Frequency-Spatial Attention Network for Medical Image Segmentation

Abstract

Medical image segmentation, a pivotal technology in computer-aided diagnosis (CAD) systems, plays a critical role in enhancing the clinical utility of lesion quantification, surgical navigation planning, and disease progression assessment. Despite its significance, the field faces persistent challenges, including the loss of high-frequency details, inadequate modeling of multi-scale structures, and limited generalization across diverse imaging modalities. To address these issues, this study introduces FSA-Net, a novel multi-domain collaborative enhancement framework for general medical image segmentation. By integrating a frequency-spatial dual-path attention mechanism with an adaptive multi-scale Transformer architecture, FSA-Net achieves robust retention of high-frequency signals, facilitates cross-organ semantic interaction, and enables precise segmentation of complex anatomical structures. Specifically, the proposed framework incorporates two key innovations: (1) a frequency-domain dynamic mask module (FSAS), which decouples high-frequency edge information from low-frequency structural components during the feature embedding phase, thereby mitigating the attenuation of high-frequency details typically caused by conventional downsampling operations; and (2) a wavelet decomposition module combined with window-based attention mechanisms, which concurrently models long-range spatial dependencies and channel-wise semantic correlations within Transformer blocks, significantly improving the framework's adaptability to cross-modal data. Extensive experiments conducted on multiple public datasets demonstrate that FSA-Net outperforms existing state-of-the-art methods in terms of segmentation accuracy and robustness. The proposed framework not only addresses the core challenges in medical image segmentation but also provides a scalable theoretical paradigm and technical foundation for accurate segmentation in complex anatomical scenarios.