Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Joint Dense-Point Representation for Contour-Aware Graph Segmentation

Kit Mills Bransby, Greg Slabaugh, Christos Bourantas, Qianni Zhang

TL;DR

This paper tackles the conflict between dense pixel-wise segmentation and contour stability in medical imaging by introducing a joint Dense-Graph framework that learns both point-based contour representations and dense pixel predictions. A two-network architecture (Dense-Graph DG and Dense-Dense DD) with Image-to-Graph Skip Connections enables multi-scale feature fusion, while six Chebyshev graph convolution layers progressively refine contour coordinates. A novel Hybrid Contour Distance (HCD) loss biases predictions toward the contour boundary via an unsigned distance field, yielding accurate, topology-stable segmentations on four Chest X-ray datasets and outperforming both dense- and graph-based baselines. The approach is efficient, generalizable to other graph segmentation tasks, and accompanied by open-source code for reproducibility and broader adoption in clinical image analysis.

Abstract

We present a novel methodology that combines graph and dense segmentation techniques by jointly learning both point and pixel contour representations, thereby leveraging the benefits of each approach. This addresses deficiencies in typical graph segmentation methods where misaligned objectives restrict the network from learning discriminative vertex and contour features. Our joint learning strategy allows for rich and diverse semantic features to be encoded, while alleviating common contour stability issues in dense-based approaches, where pixel-level objectives can lead to anatomically implausible topologies. In addition, we identify scenarios where correct predictions that fall on the contour boundary are penalised and address this with a novel hybrid contour distance loss. Our approach is validated on several Chest X-ray datasets, demonstrating clear improvements in segmentation stability and accuracy against a variety of dense- and point-based methods. Our source code is freely available at: www.github.com/kitbransby/Joint_Graph_Segmentation

Joint Dense-Point Representation for Contour-Aware Graph Segmentation

TL;DR

This paper tackles the conflict between dense pixel-wise segmentation and contour stability in medical imaging by introducing a joint Dense-Graph framework that learns both point-based contour representations and dense pixel predictions. A two-network architecture (Dense-Graph DG and Dense-Dense DD) with Image-to-Graph Skip Connections enables multi-scale feature fusion, while six Chebyshev graph convolution layers progressively refine contour coordinates. A novel Hybrid Contour Distance (HCD) loss biases predictions toward the contour boundary via an unsigned distance field, yielding accurate, topology-stable segmentations on four Chest X-ray datasets and outperforming both dense- and graph-based baselines. The approach is efficient, generalizable to other graph segmentation tasks, and accompanied by open-source code for reproducibility and broader adoption in clinical image analysis.

Abstract

We present a novel methodology that combines graph and dense segmentation techniques by jointly learning both point and pixel contour representations, thereby leveraging the benefits of each approach. This addresses deficiencies in typical graph segmentation methods where misaligned objectives restrict the network from learning discriminative vertex and contour features. Our joint learning strategy allows for rich and diverse semantic features to be encoded, while alleviating common contour stability issues in dense-based approaches, where pixel-level objectives can lead to anatomically implausible topologies. In addition, we identify scenarios where correct predictions that fall on the contour boundary are penalised and address this with a novel hybrid contour distance loss. Our approach is validated on several Chest X-ray datasets, demonstrating clear improvements in segmentation stability and accuracy against a variety of dense- and point-based methods. Our source code is freely available at: www.github.com/kitbransby/Joint_Graph_Segmentation
Paper Structure (12 sections, 2 equations, 4 figures, 2 tables)

This paper contains 12 sections, 2 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. Network Design
  4. Graph Convolutional Network
  5. Joint Dense-Point Learning
  6. Hybrid Contour Distance
  7. Experiments and Results
  8. Datasets
  9. Model Implementation & Training
  10. Comparison to Existing Methods & Ablation Study
  11. Conclusion
  12. Acknowledgements

Figures (4)

  • Figure 1: Network Architecture: a Dense-Dense network (top) enriches image features in a Dense-Graph network (bottom).
  • Figure 2: Feature map activation comparison between UNet encoder, UNet decoder, HybridGNet encoder and our encoder, using two examples. Top four most activated channels are summed channel-wise for convolutional layers 1-5 in each encoder/decoder. L$\rightarrow$R: decreasing resolution, increasing channel depth. Note, activations in our encoder consistently highlight areas which are more pertinent to segmentation
  • Figure 3: Our Hybrid Contour Distance loss biases the distance field to contours rather than the points representing the contour. Top L$\rightarrow$R: Segmentation mask represented with edges, unsigned distance field for lungs, and heart. Bottom: Effect of beta in HCD.
  • Figure 4: JSRT & Padchest: Qualitative Analysis. Note that our method does not suffer from the topological errors of dense-based methods but benefits from their segmentation accuracy. Specifically, improvements (white boxes) are most prevalent in areas of complexity such as where the heart and lungs intersect.