Research map / Graph learning / Graph self-supervised learning
GraphMorph: Tubular Structure Extraction by Morphing Predicted Graphs
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes GraphMorph, which predicts a branch-level graph of a tubular structure with a Deformable DETR-based Graph Decoder and morphs it into a topologically accurate centerline mask via a training-free SkeletonDijkstra Morph Module, plus a post-processing step that reduces segmentation false positives.
Problem
Pixel-level segmentation of tubular structures (vessels, roads) often yields broken branches, redundant branches, and topological errors, because it does not explicitly model branch-level topology.
Approach
Predict nodes (junctions and endpoints) with a modified Deformable DETR and edges with a dynamic condition-linear-layer link predictor, then find minimum-cost single-pixel-width paths between connected nodes on the centerline probability map so the output centerline mask matches the predicted graph's topology; the graph decoder also serves as auxiliary training supervision.
Key results
- On DRIVE with UNet, the Graph Decoder plus Morph Module reduced beta0 error from 2.169 (softDice) to 0.555 for centerline extraction.
- Adding GraphMorph ('softDice+Ours') to UNet on DRIVE raised segmentation Dice from 0.8148 to 0.8238 and reduced beta0 error from 1.191 to 0.692.
- On MassRoad with UNet, segmentation Dice improved from 0.7808 to 0.7849 and beta0 error dropped from 0.479 to 0.386 relative to softDice.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the Graph Decoder during training (ablation) | UNet with softDice for centerline extraction · DRIVE | β0 error: 2.169 ± 0.112 -> 1.552 ± 0.094 | Helped |
| Add the Morph Module at inference (ablation) | UNet with softDice + Graph Decoder for centerline extraction · DRIVE | β0 error: 1.552 ± 0.094 -> 0.555 ± 0.038 | Helped |
| Add the Morph Module at inference (ablation) | UNet with softDice + Graph Decoder for centerline extraction · DRIVE | Dice: 0.7506 ± 0.0127 -> 0.7496 ± 0.0118 | Hurt |
| Add GraphMorph and segmentation post-processing | UNet with softDice · DRIVE | β0 error: 1.191 ± 0.069 -> 0.692 ± 0.047 | Helped |
| Add GraphMorph and segmentation post-processing | UNet with softDice · ISBI12 | β0 error: 0.569 ± 0.046 -> 0.361 ± 0.034 | Helped |
| Add GraphMorph and segmentation post-processing | UNet with softDice · STARE | Dice: 0.8170 ± 0.0402 -> 0.8210 ± 0.0464 | Mixed |
| Add GraphMorph and segmentation post-processing | UNet with softDice · MassRoad | β0 error: 0.479 ± 0.020 -> 0.386 ± 0.016 | Helped |
| Add GraphMorph and segmentation post-processing | 3D UNet with softDice · PARSE | Dice: 0.7968 ± 0.0166 -> 0.8196 ± 0.0138 | Helped |
| Use the Dynamic link prediction module instead of the [rln]-token | GraphMorph · DRIVE centerline extraction | inference time (s): 0.2823 -> 0.1582 | Helped |
| Use the Dynamic link prediction module instead of the [rln]-token | GraphMorph · STARE centerline extraction | inference time (s): 0.1385 -> 0.0754 | Helped |
| Set ROI size to H=32 rather than reducing it (ablation) | GraphMorph on DRIVE · DRIVE centerline extraction | β0 error | Helped |
| Set the Morph Module path threshold to p_thresh = 0.5 (ablation) | GraphMorph centerline extraction · Centerline extraction | topological metrics | Helped |
Surprising findings
- Adding the Morph Module at inference slightly lowered volumetric metrics (e.g., Dice on DRIVE 0.7506 to 0.7496) while strongly improving topological metrics.
- The Graph Decoder adds substantial training cost: GPU memory rose from 5.4 GB to 11.8 GB and time per iteration from 0.203 s to 0.589 s.
Limitations
- Segmentation still relies on a post-processing step rather than using predicted graphs directly.
- Sliding-window inference over small ROIs may miss global context and comprehensive branch-level detail.
- The Morph Module's path threshold can miss true-positive edges when the probability map is inaccurate; 3D results are preliminary on a single dataset.
Experimental setup
Datasets: DRIVE STARE ISBI12 MassRoad PARSE
Benchmarks: DRIVE STARE ISBI12 MassRoad PARSE
Metrics: Dice clDice AUC ACC beta0 error beta1 error chi error ARI VOI
Baselines: softDice clDice PointScatter TopoLoss DSCNet TransUNet DC-UNet LinkNet34 D-Linknet34
Builds on: Deformable DETR Dijkstra algorithm clDice PointScatter RelationFormer
Closest papers by meaning
- clDice - A Novel Topology-Preserving Loss Function for Tubular Structure Segmentation (CVPR 2021 · similarity 0.88)
- Deep Distance Transform for Tubular Structure Segmentation in CT Scans (CVPR 2020 · similarity 0.86)
- Topograph: An Efficient Graph-Based Framework for Strictly Topology Preserving Image Segmentation (ICLR 2025 · similarity 0.84)
- SegGraph: Leveraging Graphs of SAM Segments for Few-Shot 3D Part Segmentation (NeurIPS 2025 · similarity 0.83)
- DeformCL: Learning Deformable Centerline Representation for Vessel Extraction in 3D Medical Image (CVPR 2025 · similarity 0.82)
- LaneGraph2Seq: Lane Topology Extraction with Language Model via Vertex-Edge Encoding and Connectivity Enhancement (AAAI 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.