DCFFSNet: Deep Connectivity Feature Fusion Separation Network for Medical Image Segmentation

Xun Ye, Ruixiang Tang, Mingda Zhang, Jianglong QinarXiv 2025 · DOI · arXiv

Abstract

Accurate medical image segmentation is critical for digital health, yet traditional deep learning models frequently produce fragmented segments and blurry edges. This paper introduces DCFFSNet, a Dual-Connectivity Feature Fusion-Separation Network designed to explicitly decouple topological connectivity from standard classification features. By utilizing a fusion-separation strategy, the model adaptively weights these distinct information types across multiple scales. Experimental results on skin lesion and nuclei datasets demonstrate that DCFFSNet achieves state-of-the-art accuracy with significantly lower computational costs than many established architectures.

Introduction

In medical imaging, identifying organs or lesions at the pixel level requires not just classification but also topological consistency. Connectivity—how adjacent pixels relate—is vital for ensuring segmented regions are spatially continuous with sharp boundaries. Existing "connectivity-based" networks often struggle because they "forcibly inject" connectivity as an auxiliary feature, leading to a coupled feature space where categorical features (like texture) and topological features interfere. DCFFSNet addresses this by rethinking the interaction between these feature types through a dedicated decoupling mechanism.

Methods

DCFFSNet employs a U-shaped encoder-decoder structure, as shown in Figure 1, integrated with four specialized modules to enhance feature representation and separation.

The overall architecture of DCFFSNet featuring the encoder-decoder structure and specialized modules.
  1. DSCRIM (Deeply Supervised Connectivity Representation Injection Module): Positioned at the bottleneck, this module uses deep supervision to prime the feature space with connectivity data, dividing features into eight groups corresponding to pixel adjacency directions.
  2. MSFFM (Multi-Scale Feature Fusion Module): This core component uses a self-attention mechanism to decouple the feature space. It dynamically calculates weights for connectivity and standard features, allowing the network to prioritize information types based on the current scale.
  3. MSRCM (Multi-Scale Residual Convolution Module): Employed during upsampling, this module uses a pyramid of convolution kernels (1x1 to 7x7) to capture multi-scale context and preserve fine details.
  4. PConv (Directional Convolution): This module optimizes the final connectivity mask by aligning feature maps with the eight neighbor directions, ensuring topological soundness in the output.

Results and Discussion

The performance of DCFFSNet was evaluated on three benchmarks: ISIC2018 (skin lesions), DSB2018 (nuclei), and MoNuSeg (multi-organ nuclei).

Quantitative Performance

As detailed in Figure 6, DCFFSNet consistently outperformed heavy-weight models such as TransUNet and CMUNet. Notably, it achieved a 90.0% Dice score on ISIC2018 and 92.1% on DSB2018.

Comparison of DCFFSNet against state-of-the-art models across three datasets, highlighting IoU, Dice scores, and computational efficiency.

Efficiency and Insights

A key advantage of DCFFSNet is its efficiency. Requiring only 21.732 GFLOPs, it is more computationally lightweight than standard U-Net (50.1G) and UNet++ (106.1G). The researchers found that decoupling features is superior to simple injection, as it prevents interference between topological and categorical data. Furthermore, explicit directional awareness leads to smoother edge transitions and eliminates fragmentation artifacts.

Conclusion

DCFFSNet represents a significant advancement in medical image segmentation by introducing a fusion-separation strategy to manage topological connectivity. While the model excels at local edge precision, future work will explore integrating global attention mechanisms, such as Transformers, to better capture large-scale structural patterns while maintaining local topological integrity.

Reference: Zhang, M., et al. "DCFFSNet: Deep Connectivity Feature Fusion Separation Network for Medical Image Segmentation." arXiv:2507.18407v2, 2025.

Summary updated

BibTeX
@misc{ye2025dcffsnet,
  author = {Xun Ye and Ruixiang Tang and Mingda Zhang and Jianglong Qin},
  title = {{DCFFSNet: Deep Connectivity Feature Fusion Separation Network for Medical Image Segmentation}},
  year = {2025},
  eprint = {2507.18407},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2507.18407},
}
Figure