FFG-Unet: A medical image segmentation model with frequency feature guidance

Deep learning-based medical image segmentation methods have shown considerable clinical potential for lesion recognition and localization. However, medical images are characterized by low lesion contrast and poorly defined boundaries, making accurate segmentation particularly challenging. To address these challenges, we propose FFG-Unet, a medical image segmentation model guided by frequency features. By integrating both high- and low-frequency features throughout the segmentation process, FFG-Unet maintains high sensitivity to lesion boundary details in both the encoding and decoding stages. In terms of the model design, we propose a High-Frequency Feature Extraction Block to capture detailed high-frequency information. In addition, a Multi-Scale High-Frequency Feature Guidance Block is introduced to incorporate high-frequency features and enhance detail and edge perception during encoding. Finally, we design a Frequency Feature Fusion Block to extract low-frequency features via wavelet transform and fuse them with high-frequency features to guide the decoding process, effectively alleviating edge blurring and incomplete lesion segmentation. We conduct extensive experiments on four public datasets, including BUSI, ISIC, Prostate, and Kvasir. The model achieves IoU scores of 70.85%, 82.38%, 87.71% and 81.68%, respectively. These results demonstrate the effectiveness and competitiveness of the proposed method for precise medical image segmentation. Our code is available at: https://github.com/LinSyyyy/FFG-Unet.