Complementary fusion network of scaling attention and global attention for volumetric medical image segmentation

Medical image segmentation is crucial in the diagnosis and treatment of diseases. Attention mechanisms have been widely adopted to highlight salient regions while suppressing irrelevant information. Although channel and spatial attention can emphasize a few key features within a limited local range, they struggle to model the global contextual dependencies. In contrast, self-attention in the transformer is capable of modeling such global relationships. Integrating these methods can therefore leverage their complementary strengths for more comprehensive feature representation.

Existing hybrid works integrate the two types of attention mechanisms in a cascading manner, which may disrupt both local key features and global contextual dependencies. This study aims to design a parallel fusion strategy that jointly exploits local and global information to achieve superior segmentation performance.

We propose a complementary fusion network (CFNet) for volumetric medical image segmentation. To learn local representations, we introduce a novel scaling attention mechanism that redefines both channel and spatial attention. The channel attention employs global average and max pooling to simultaneously capture tissue texture and fine-grained responses. Adaptive convolution is then introduced to efficiently facilitate channel interaction within a local receptive field. The spatial attention uses atrous convolution to enlarge the receptive field and capture rich spatial details. To jointly model both local and global dependencies, we design a parallel mixed module consisting of the proposed scaling attention and transformer-based global attention, achieving continuous and complementary feature learning.

We comprehensively evaluated CFNet on four benchmark segmentation tasks, including abdominal multi-organ, cardiac, brain tumor, and left atrium segmentation. Our method achieved Dice coefficients of 87.15%, 92.31%, 86.4%, and 93.91% for the respective tasks.

Experimental results demonstrate that our method outperforms state-of-the-art methods. These superior results highlight the potential of CFNet to support clinical decision-making and treatment planning.