In recent years, Transformer architectures have gained prominence in the domain of medical image segmentation due to their robust capabilities in modeling long-range dependencies. Nonetheless, these methods exhibit notable limitations in local feature extraction, posing challenges for existing U-shaped segmentation networks in achieving a balance between global context awareness and local detail preservation, particularly in the complex task of segmenting organ boundaries. To address this issue, this study introduces CE-SwinUnet, a lightweight segmentation model that achieves a performance breakthrough through a dual-module synergy mechanism. The incorporation of a gated cross-attention module within the skip connection allows for dynamic adjustment of feature fusion weights between the encoder and decoder, significantly enhancing the reuse rate of effective features and effectively resolving the issue of multi-organ boundary ambiguity. In addition, a channel-efficient complementary module is integrated at the encoder, decoder, and bottleneck layer to synchronize the capture of weighted global context and local channel dependencies through the synergistic mechanism of efficient attention and channel attention. The efficacy of the proposed method was validated on the Synapse dataset, with the CE-SwinUnet model achieving 81.77% and 19.63 mm in Dice and Hausdorff distance metrics, respectively, surpassing existing medical image segmentation models. The CE-SwinUnet network addresses the challenges of long-distance-dependent modeling and localized feature extraction while maintaining lightweight characteristics, thereby significantly enhancing boundary segmentation accuracy.
Sun et al. (Sun,) studied this question.