Structure-Preserving Instance Segmentation via Skeleton-Aware Distance Transform

Zudi Lin, Donglai Wei, Aarush Gupta, Xingyu Liu, Deqing Sun, Hanspeter Pfister

Abstract

Objects with complex structures pose significant challenges to existing instance segmentation methods that rely on boundary or affinity maps, which are vulnerable to small errors around contacting pixels that cause noticeable connectivity change. While the distance transform (DT) makes instance interiors and boundaries more distinguishable, it tends to overlook the intra-object connectivity for instances with varying width and result in over-segmentation. To address these challenges, we propose a skeleton-aware distance transform (SDT) that combines the merits of object skeleton in preserving connectivity and DT in modeling geometric arrangement to represent instances with arbitrary structures. Comprehensive experiments on histopathology image segmentation demonstrate that SDT achieves state-of-the-art performance.

Introduction

In biomedical imaging, segmenting individual objects (instances) is a complex task due to non-convex shapes, varying widths, and high object density. Traditional bottom-up methods typically utilize boundary maps or standard distance transforms (DT). Boundary maps are highly sensitive to noise; a single-pixel error can lead to the merging of adjacent instances. Conversely, standard DT, which maps pixels to their distance from the nearest boundary, often produces multiple local maxima within a single complex object, leading to over-segmentation.

Comparison of instance representations: SDT preserves connectivity where standard DT fails.

As shown in Figure 1, standard DT (c) fails to maintain connectivity in narrow regions (indicated by the red arrow), whereas the proposed Skeleton-aware DT (d) preserves the global structure of the instance by leveraging the object's skeleton.

Methodology

The core innovation of this work is the Skeleton-Aware Distance Transform (SDT), which fuses the topological stability of object skeletons with the geometric precision of distance transforms.

The SDT Energy Function

The SDT defines an energy landscape $E_{SDT}$ for each pixel $p$ within an instance $\Omega$. It considers both the distance to the boundary ($d_b$) and the distance to the object's skeleton ($d_s$): $$E_{SDT}(p) = \left( \frac{d_b}{d_s + d_b} \right)^\alpha$$ This formulation ensures that the energy is exactly 1 on the skeleton and 0 at the boundary, creating a single global maximum for each instance regardless of its geometric complexity. The parameter $\alpha$ (typically set to 0.8) controls the curvature of the energy landscape to facilitate network learning.

System Architecture

The pipeline consists of a training phase and an inference phase. A Fully Convolutional Network (FCN), specifically DeepLabV3 with a ResNet backbone, is trained to predict the SDT map. The authors found that quantizing the energy range into discrete bins and treating the prediction as a classification task improved robustness over direct regression.

Overview of the training and inference pipeline using SDT.

During inference, the predicted SDT map is thresholded to extract skeleton "seeds." These seeds are then expanded into full instance masks using a watershed algorithm, ensuring that each seed corresponds to a unique object and preventing over-segmentation.

Experiments and Results

The method was evaluated on the Gland Segmentation Challenge (GlaS) dataset, consisting of 165 microscopy images of colon histology. This dataset is particularly challenging due to the irregular and touching nature of the glands.

SDT achieved state-of-the-art results, outperforming existing methods in 5 out of 6 evaluation metrics. Notably, it achieved a 10.6% improvement in Hausdorff distance, indicating significantly higher shape accuracy. Qualitative analysis confirms that SDT successfully maintains the connectivity of elongated structures while preventing the merging of adjacent glands.

Conclusion

The Skeleton-Aware Distance Transform provides a robust representation for instance segmentation of complex structures. By integrating skeletal topology, it overcomes the over-segmentation issues inherent in standard distance transforms. Future work includes extending this framework to 3D medical imaging and refining skeleton generation to handle noisier boundaries.

Summary updated

Abstract

Objects with complex structures pose significant challenges to existing instance segmentation methods that rely on boundary or affinity maps, which are vulnerable to small errors around contacting pixels that cause noticeable connectivity change. While the distance transform (DT) makes instance interiors and boundaries more distinguishable, it tends to overlook the intra-object connectivity for instances with varying width and result in over-segmentation. To address these challenges, we propose a skeleton-aware distance transform (SDT) that combines the merits of object skeleton in preserving connectivity and DT in modeling geometric arrangement to represent instances with arbitrary structures. Comprehensive experiments on histopathology image segmentation demonstrate that SDT achieves state-of-the-art performance.
BibTeX
@inproceedings{lin2023structure,
  author = {Zudi Lin and Donglai Wei and Aarush Gupta and Xingyu Liu and Deqing Sun and Hanspeter Pfister},
  title = {{Structure-Preserving Instance Segmentation via Skeleton-Aware Distance Transform}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2023},
  url = {https://openreview.net/forum?id=67GMVv2Mtn},
}
Figure