MSFFE-Net: Multi-scale Spatial-Frequency Feature Enhancement for accurate liver tumor segmentation

Abstract

Accurate liver tumor segmentation is a crucial aspect for early diagnosis and surgery, but existing segmentation methods struggle with tumor heterogeneity, unclear boundaries, and small lesions due to limited multi-scale feature fusion and spatial perception. To alleviate these issues, we propose MSFFE-Net, a novel segmentation network that imposes a Multi-scale Spatial-Frequency Feature Enhancement mechanism, with the objective of unifying spatial and frequency domains to enrich feature representational power. Moreover, a Spatial-Frequency Domain Fusion (SFDF) module is incorporated to unify Fourier features with a dual-branch encoder, where standard convolutions and Residual Dilated Convolutions (RDC) are jointly employed to enable multi-scale feature extraction and to enhance edge perception. In addition, a Multi-scale Semantic Enhancement (MSE) module is introduced at the bottleneck to model global context, and CBAM attention is integrated into the skip connections to further optimize feature aggregation. Extensive experiments on the LiTS_2017 and 3Dircadb datasets further validate the effectiveness of the proposed method, achieving Dice coefficients of 98.12% and 97.24% for liver segmentation, and 89.61% and 92.87% for tumor segmentation, respectively. Compared with mainstream approaches such as nnU-Net and TransUNet, our model delivers Dice gains of 0.07%, 2.57%, and 1.00%, 1.83% on complex tumor datasets. In addition, the architecture maintains a favorable trade-off between accuracy and efficiency, with only 17.46 MB of parameters and an inference speed of 68.74 FPS. Ablation studies validate the model’s effectiveness in complex boundary and small target segmentation, advancing intelligent liver cancer diagnosis with potential for other organs tumor segmentation.