Directional Connectivity-based Segmentation of Medical Images

Ziyun Yang, Sina Farsiu

Abstract

Standard deep learning approaches to medical image segmentation often treat the task as a per-pixel classification problem, frequently resulting in topological inconsistencies such as fragmented vessels or anatomically implausible boundaries. We present DconnNet, a novel framework that incorporates directional connectivity to enhance segmentation accuracy and topological integrity. By disentangling directional information from categorical features through a Sub-path Direction Excitation (SDE) module and utilizing an Interactive Feature-space Decoder (IFD), DconnNet effectively models spatial relationships. Furthermore, a Size Density Loss (SDL) is introduced to mitigate class imbalance. Experimental results across three benchmarks—Retouch, ISIC2018, and CHASEDB1—demonstrate that DconnNet significantly outperforms state-of-the-art methods in both volumetric and topological metrics.

Introduction

In medical imaging, the structural shape and connectivity of an object are often as clinically significant as its identification. Standard segmentation models typically classify pixels independently, which ignores the inherent spatial relationships between neighboring pixels. This lack of spatial awareness often leads to "broken" results, such as disconnected blood vessels or fluid pockets with jagged, anatomically impossible boundaries.

While previous research has attempted to use connectivity masks to describe how a pixel relates to its eight neighbors, these methods often lump all information into a single processing path. This "entangled" latent space makes it difficult for a network to distinguish between categorical features (what an object is) and directional features (how it is oriented or connected). DconnNet addresses these limitations by explicitly modeling directional connectivity as a topological prior.

Methods

DconnNet introduces three primary innovations designed to improve spatial reasoning and handle the unique challenges of medical datasets:

1. Sub-path Direction Excitation (SDE) Module

To prevent the entanglement of categorical and spatial data, the SDE module "slices" the latent space into eight sub-paths, each corresponding to a specific neighbor direction. By processing these directions independently, the module extracts a "directional prior" that guides the network. This disentanglement allows the model to learn the rules of digital topology—such as separation and connectedness—more effectively than standard architectures.

2. Interactive Feature-space Decoder (IFD)

The IFD utilizes two parallel flows: space flow and feature flow. These flows interact dynamically across different scales, allowing high-level directional embeddings to refine the main feature maps in a coarse-to-fine manner. This dual-flow approach ensures that the high-level understanding of connectivity informs the low-level pixel classification.

3. Size Density Loss (SDL)

Medical datasets are notoriously imbalanced, often featuring tiny biomarkers (like fluid leaks) within massive scans. SDL is a novel loss function that weights pixel importance based on the label size distribution of the entire dataset. This helps the model focus on small but critical structures that might otherwise be ignored by standard loss functions like Cross-Entropy or Dice loss.

Experimental Results

The model was evaluated on three diverse public benchmarks to test its versatility and topological accuracy:

  • Retouch (Retinal Fluid): DconnNet achieved a Volume Dice ($DSC_v$) of 78.2%, a significant improvement over the next best competitor, CPFNet (69.0%). The model was particularly effective at identifying tiny fluid regions missed by other models.
  • ISIC2018 (Skin Lesions): The model reached a Dice score of 90.43%, surpassing specialized models like Ms RED.
  • CHASEDB1 (Retinal Vessels): This dataset specifically tested the model's ability to maintain thin, connected structures. DconnNet excelled in topological metrics, achieving a clDice of 83.3% and superior Betti numbers, indicating that the predicted vessel networks were more anatomically realistic and better connected than those produced by standard methods.

Discussion and Conclusion

Analysis of the model's latent space via T-SNE visualizations confirms that the SDE module successfully groups directional features into distinct clusters, proving that the network learns to separate "direction" from "category." Furthermore, the channel-wise slicing in the SDE module improves efficiency; by reducing the dimensionality of each sub-path, the model reduces the parameter count for attention mechanisms compared to standard dual-attention units.

While DconnNet performs exceptionally well on 2D slices, future work will focus on extending these principles to 3D connectivity modeling, as medical data is inherently volumetric. Additionally, while the Size Density Loss is highly effective, it requires pre-calculating label distributions, which may require adjustment when applying the model to new clinical populations. Overall, DconnNet represents a significant step forward in integrating topological priors into deep learning for medical image analysis.

Summary updated

BibTeX
@inproceedings{yang2023directional,
  author = {Ziyun Yang and Sina Farsiu},
  title = {{Directional Connectivity-based Segmentation of Medical Images}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2023},
  doi = {10.1109/cvpr52729.2023.01109},
}
Figure