S-Net: A novel shallow network for enhanced detail retention in medical image segmentation.

BACKGROUND AND OBJECTIVE

In recent years, deep U-shaped network architectures have been widely applied to medical image segmentation tasks, achieving notable successes. However, the inherent limitation of this architecture is that multiple down-sampling lead to significant loss of input image detail information. A series of improvements in skip connections designed to enhance information transfer have not fundamentally resolved the issue. Therefore, we consider retaining information in a simpler and more effective way.

METHODS

In this paper, we propose a novel shallow network, S-Net, which contains only two output resolution stages, allowing for the preservation of more detailed information from the input images. To address the challenge of shallow networks primarily relying on high-resolution feature maps as the main information flow, we propose a Global-Local Feature Fusion (GLFF) module at the network bottleneck layer. This module integrates the superior global contextual information extraction capabilities of Mamba with the local feature capturing abilities of multi-scale depthwise convolutions, enabling the extraction of crucial semantic features from high-resolution feature maps within a shallow network architecture, while maintaining a smaller model size.

RESULTS

Extensive experiments on four different types of medical image datasets show that S-Net achieves the best segmentation performance compared to existing models, with more refined segmentation details. For example, on ultrasound datasets (BUSI), the IOU is 2.95% higher and DICE is 2.27% higher than the second-best model. Additionally, S-Net has only 1.52M parameters, making it competitive in terms of lightweight design.

CONCLUSIONS

Comparative and ablation experiments demonstrate the efficiency of the proposed architecture and modules. It shows that we do not need many down-sampling operations to reduce the size of feature maps significantly. This work provides new research ideas for further improving the accuracy of medical image segmentation and expands the research direction for model lightweight design. The code will be available at: https://github.com/qinghua0715/S-Net.