Source-linked AI summary
Temporally Consistent Graph Extraction and Matching for Longitudinal Angiographic Images
Linus Kreitner, Laurin Lux, Carmen Baumann, Daniel Rueckert, Martin J. Menten
TL;DR
Vessel-graph pipelines can resolve vessel-level temporal changes, but small segmentation differences make sequential graph extraction, refinement, and matching highly sensitive. The paper matches graphs before jointly refining them, and reports higher matched area without graph fragmentation than separate or no refinement in complex retinal vessel graphs.
Problem
Small variations in angiographic segmentation maps can produce substantially different vessel graphs, inhibiting accurate matching of sequential graphs from the same subject.
Method
The method performs early matching after basic graph extraction, then jointly refines both graphs by removing spurious bulges and merging junctions.
Results
The strategy produces higher matched area without graph fragmentation than separate or no refinement in complex retinal vessel graphs.
Takeaways & Limitations
Early matching before joint refinement supports longitudinal vessel-graph correspondence in the evaluated complex retinal setting.
Takeaways & Limitations
The results were obtained in a carefully crafted setting using a specifically developed graph extraction and matching algorithm.
Abstract
from arXiv · showhide
Recent advances in angiographic imaging have enabled longitudinal visualization of the microvasculature. Image processing pipelines based on vessel graphs are able to resolve subtle temporal changes at the level of individual blood vessels. However, current strategies for graph extraction, refinement, and matching are highly sensitive, with even minuscule differences in the underlying segmentation map resulting in substantially different vessel graphs. These artifacts severely inhibit the ability to accurately match sequential vessel graphs of the same subject over time. To address this problem, we propose a strategy that matches graphs before jointly refining them. Specifically, we perform an early matching after basic graph extraction before removing spurious bulges and merging junctions in both graphs using joint information. In experiments with complex retinal vessel graphs, we demonstrate that this strategy results in a higher matched area without graph fragmentation compared to separate or no refinement, respectively.
1 Introduction
Longitudinal angiographic analysis seeks vessel-level temporal biomarkers, but small segmentation differences can destabilize graph construction and matching. The paper addresses this by matching graphs early and then jointly refining them, demonstrating utility on challenging retinal graphs.
- 1 Introduction: Vessel graphs encode bifurcations and connecting segments as nodes and edges, optionally enriched with thickness, length, curvature, or branching-angle features.These representations support fine-grained analysis of temporal changes at individual blood vessels.
- 1 Introduction: Small segmentation variations can produce vastly different vessel graphs and disproportionately affect graph refinement.This sensitivity inhibits reliable correspondence between sequential graphs, even when the underlying vasculature remains similar.
- 1 Introduction: The proposed strategy matches graphs after basic extraction, before jointly removing spurious bulges and merging junctions using information from both graphs.This ordering is intended to reduce the instability introduced by independent refinement.
- 1 Introduction: The method is evaluated on a unique OCTA dataset with repeated intra-subject scans and complex longitudinal vessel graphs containing 2,000 edges.The evaluation targets a challenging setting for matching sequential vessel graphs from the same subject.
2 Methods
The method matches raw longitudinal vessel graphs before jointly refining them, coupling artifact removal and junction merging across time points to improve correspondence.
- 2.4 Joint refinement of matched graphs: Early matching precedes bulge removal and junction merging, allowing both graphs to be refined using joint information.The pipeline extracts graphs separately, matches them in their raw form, then performs coupled refinement.
- 2.1 Graph extraction: Graph extraction converts segmentation skeletons into nodes and edges by grouping adjacent junctions into hubs and assigning shortest paths between hubs.Node and edge radii are also computed from distance-transform measurements.
- 2.2 Graph refinement: Separate refinement removes spurious bulges below a threshold and merges degree-3 junction pairs connected by short H-shaped bridges.The merging cost penalizes long, thin bridges and evaluates geometric compatibility of incident edges.
- 2.3 Graph matching: Flexible path matching relaxes strict edge bijection by mapping concatenated edge paths, accommodating splintered edges caused by intermediate bulges.Candidate costs combine distance, rotation, shape, area, endpoint, and intermediate-junction terms before optimization.
- 2.3 Graph matching: Candidate matches are selected with binary decision variables that enforce at most one selected candidate per edge while penalizing unmatched edges.The resulting path matches are decomposed into direct one-to-one edge correspondences by inserting degree-2 nodes where needed.
- 2.4 Joint refinement of matched graphs: Joint refinement preserves a structure when both its bulge or H-bridge criteria are satisfied across matched graphs, while unmatched edges follow the original rules.This prevents one graph from independently removing a small structure that remains present at another time point.
3 Experiments and Results
Experiments on repeated retinal OCTA scans compare unrefined, separately refined, and jointly refined graph construction, alongside established tools. Early matching with joint refinement increases matched area while preserving stable or improved matching quality and reducing fragmentation.
- Dataset: The dataset contains 160 retinal OCTA images from two repeated scans of 80 healthy eyes acquired on the same day.Each scan uses a 1.5 × 1.5 mm2 en-face projection with a detailed 608×608 pixel segmentation mask.
- Graph refinement: About 1,300 edges occur on average in unrefined graphs, mostly as very small segments caused by segmentation variations along vessel boundaries.Bulge removal and junction merging reduce fragmentation and produce a more coherent graph topology.
- Experimental design: The comparison includes no refinement, separate refinement, joint refinement, and the established Voreen and VesselVio graph extraction tools.The experiments use repeated scans with expected minor structural changes, isolating artifacts in graph construction.
- Matching quality: Representative examples show that no refinement fragments edges, separate refinement yields unmatched or wrongly paired edges, and joint refinement produces convincing edge-level matches.These qualitative differences complement the quantitative comparison of matched area and edge matching quality.
- Results: Joint refinement increases the total matched area while maintaining stable match quality compared with separate refinement.The baseline produces more matches, but its biomarker correspondence between matched vessels is substantially lower.
4 Discussion
The discussion identifies sensitivity to small image variations as a major obstacle to matching longitudinal vessel graphs and presents early matching before joint refinement as the remedy. The authors report higher matched area without fragmentation in complex retinal graphs, while noting that results were obtained in a carefully crafted setting.
- 4 Discussion: Small variations in input images can substantially alter vessel graphs and severely inhibit matching even when the underlying vasculature corresponds.This sensitivity is especially relevant to longitudinal analysis of individual blood vessels.
- 4 Discussion: Early graph matching before joint refinement yields higher matched area without graph fragmentation than separate or no refinement in complex retinal vessel graphs.The strategy matches graphs before removing spurious bulges and merging junctions using joint information.
- 4 Discussion: The reported results come from a carefully crafted setting using a specifically developed graph extraction and matching algorithm.The authors nevertheless state that the strategy is conceptually compatible with other graph-construction methods, three-dimensional images, and more than two vessel graphs.