Defining Functional Liver Segments Via Vessel Structure: A Novel Graph Neural Network (GNN) Framework for Anatomically Consistent Couinaud Segmentation on CT
Abstract
Purpose
To propose a Graph Neural Network (GNN) framework that explicitly models liver vessel topology to detect anatomical landmarks and reconstruct anatomically consistent segments.
Methods
Using 130 contrast-enhanced CT cases (100 training, 30 testing) from the Medical Segmentation Decathlon Task08 dataset, liver and vessel masks were first obtained using a previously developed segmentation module. Vessel centerlines were extracted and sampled to construct patient-specific undirected graphs, with nodes representing vessel points and edges encoding local topology. A graph attention neural network was trained in a supervised manner to classify graph nodes into seven clinically defined vascular landmarks, including superior and inferior vena cava points, middle, right, and left hepatic veins, and right and left portal vein landmarks. Landmark ground truth was derived from Couinaud annotations by projecting segment boundaries onto vessel centerlines. Predicted vessel landmarks were aggregated and used in an analytic, plane-based reconstruction scheme to generate Couinaud segments without training an additional reconstruction network. Performance was evaluated against Swin-UNetR (voxel-based) and a prior landmark-based method using Mean Euclidean Distance (ED), Dice Similarity Coefficient (DSC), and Average Surface Distance (ASD). Statistical significance was assessed using two-tailed Wilcoxon signed-rank tests for all paired comparisons.
Results
The GNN framework significantly outperformed baselines in landmark localization (ED: 10.89±3.4mm, p<0.02) and segmentation overlap (DSC: 82.71±2.9%, p<0.02). While ASD (2.47±0.7mm) showed moderate performance compared to landmark-based method, qualitative results showed that our method eliminates "floating islands" and fragmented segments typical of voxel-based approaches like Swin-UNetR while providing more precise plane-based boundaries than the prior landmark-based method.
Conclusion
Explicit graph-based modeling of liver vasculature enables accurate detection of functional anatomical landmarks and robust Couinaud segmentation. This framework offers a principled and interpretable alternative to purely voxel-based methods and shows promise for clinically reliable liver segmentation.