Enhancing Boundary Segmentation for Topological Accuracy with Skeleton-based Methods

Chuni Liu, Boyuan Ma, Xiaojuan Ban, Yujie Xie, Hao Wang, Weihua Xue, Jingchao Ma, Ke XuIJCAI 2024

Abstract

Accurate boundary segmentation in reticular structures—such as neural networks, road maps, and metallic grains—is critical for maintaining topological consistency. Standard pixel-wise losses, like Dice or Cross-Entropy, often fail to prevent fractures or incorrect closures that drastically alter connectivity, even when achieving high pixel-level accuracy. This paper introduces the Skea-Topo Aware loss, a novel training objective comprising a Skeleton-aware Weighted Loss (Skeaw) and a Boundary Rectified Term (BoRT). By leveraging morphological skeletons to identify and penalize topologically critical pixels, the proposed method significantly improves topological accuracy across biological, material science, and remote sensing datasets, outperforming 13 state-of-the-art approaches and improving topological consistency by up to 7 points.

Introduction

In many computer vision tasks, particularly those involving "reticular" (net-like) structures, pixel-level accuracy metrics like the Dice score can be misleading. A model might achieve high accuracy but still fail to close a cell boundary or accidentally fracture a road segment. These small errors lead to catastrophic topological changes, altering the number of connected components and the overall layout of the network. For applications like brain circuit mapping or autonomous navigation, a broken connection is far more damaging than a slightly misaligned boundary. Existing topology-aware losses often rely on computationally expensive methods like persistent homology or iterative warping, which struggle with efficiency and irregular shapes.

Methodology: Skea-Topo Aware Loss

The authors propose the Skea-Topo Aware loss to address these challenges through two primary components: the Skeleton-aware Weighted Loss (Skeaw) and the Boundary Rectified Term (BoRT).

Overview of the Skea-Topo pipeline, including the calculation of topological critical pixels and the integration of Skeaw and BoRT losses.

Skeleton-aware Weighted Loss (Skeaw)

Traditional distance-based weighting often neglects narrow or irregular regions. Skeaw utilizes the morphological skeleton of objects to normalize weights, ensuring that fragile, narrow parts of a structure receive sufficient attention during training. This geometric modeling helps prevent fractures and incorrect closures in complex shapes.

Boundary Rectified Term (BoRT)

BoRT is a computationally efficient ($O(n)$) "plug-and-play" term designed to identify topological critical pixels. These are specific pixels that, if misclassified, directly change the topology of the segmentation. Instead of complex mathematical machinery, BoRT identifies these pixels by analyzing the intersection of foreground and background skeletons with error regions. By applying a heavy penalty to these critical areas, the model learns to prioritize topological integrity.

Experimental Evaluation

The method was evaluated across three distinct domains: electron microscopy images of neurites (SNEMI3D), microscopic slices of iron grain boundaries (IRON), and aerial imagery of road networks (MASS. ROAD).

Key Results

The Skea-Topo method demonstrated superior performance across all datasets. On the SNEMI3D dataset, it improved the Variation of Information (VI)—a key metric for topological consistency—by up to 7 points compared to 13 state-of-the-art methods, including clDice and Warp loss. It also consistently achieved the lowest Betti Error, directly indicating a more accurate count of holes and connected components.

Qualitative comparison of segmentation results on the SNEMI3D dataset. Skea-Topo successfully maintains connectivity in neurite boundaries where other methods fail.

Qualitative analysis confirms that Skea-Topo effectively closes gaps in cell membranes and prevents the "ghost" road segments frequently produced by other models. Furthermore, the authors showed that BoRT can be easily integrated into standard loss functions to enhance their topological performance.

Conclusion and Future Work

The Skea-Topo Aware loss provides a robust and efficient solution for maintaining topological accuracy in boundary segmentation. By using skeletons as proxies for topology, the method avoids the computational overhead of previous approaches while delivering state-of-the-art results.

Future research directions include extending the 2D skeletonization logic to 3D volumes and adapting the framework for multi-class segmentation tasks. Additionally, further investigation into hyperparameter tuning could streamline the application of this method to new datasets.

Summary updated

BibTeX
@inproceedings{liu2024enhancing,
  author = {Chuni Liu and Boyuan Ma and Xiaojuan Ban and Yujie Xie and Hao Wang and Weihua Xue and Jingchao Ma and Ke Xu},
  title = {{Enhancing Boundary Segmentation for Topological Accuracy with Skeleton-based Methods}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2024},
}
Figure