VasGuideNet: Vascular Topology-Guided Couinaud Liver Segmentation with Structural Contrastive Loss

Abstract: Accurate Couinaud liver segmentation is critical for preoperative surgical planning and tumor localization.However, existing methods primarily rely on image intensity and spatial location cues, without explicitly modeling vascular topology. As a result, they often produce indistinct boundaries near vessels and show limited generalization under anatomical variability.We propose VasGuideNet, the first Couinaud segmentation framework explicitly guided by vascular topology. Specifically, skeletonized vessels, Euclidean distance transform (EDT)--derived geometry, and k-nearest neighbor (kNN) connectivity are encoded into topology features using Graph Convolutional Networks (GCNs). These features are then injected into a 3D encoder--decoder backbone via a cross-attention fusion module. To further improve inter-class separability and anatomical consistency, we introduce a Structural Contrastive Loss (SCL) with a global memory bank.On Task08_HepaticVessel and our private LASSD dataset, VasGuideNet achieves Dice scores of 83.68% and 76.65% with RVDs of 1.68 and 7.08, respectively. It consistently outperforms representative baselines including UNETR, Swin UNETR, and G-UNETR++, delivering higher Dice/mIoU and lower RVD across datasets, demonstrating its effectiveness for anatomically consistent segmentation. Code is available at https://github.com/Qacket/VasGuideNet.git.

• The paper introduces VasGuideNet, which encodes vessel skeletons, radii, orientations, and connectivity through GCN cross-attention to guide 3D Transformer-based Couinaud liver segmentation.
• The method achieves 83.68% Dice and 1.68% relative volume difference on Task08_HepaticVessel, plus 76.65% Dice and 7.08% relative volume difference on the private LASSD cohort.
• Structural Contrastive Loss and confidence-aware memory updates improve anatomical feature separation and volume stability, although the approach requires expert vessel masks at inference time.

VasGuideNet addresses Couinaud liver segmentation by making vascular topology an explicit first-class input rather than an implicit regularizer. The framework assumes expert-annotated vessel masks are available at both training and inference time, and channels skeleton-, distance-, and connectivity-derived vessel geometry into a 3D hierarchical Transformer backbone through a GCN-based cross-attention module. A complementary Structural Contrastive Loss (SCL) with a confidence-aware memory bank enforces anatomically consistent feature separation. On Task08_HepaticVessel and a private LASSD cohort, the method achieves Dice scores of 83.68% and 76.65% with relative volume differences of 1.68 and 7.08, respectively, outperforming UNETR, SegFormer, SegMamba, G-UNETR++, LiverFormer, and 3D-HCFormer.

Motivation and problem setting

Couinaud segmentation partitions the liver into eight functionally independent segments defined by portal and hepatic vein distribution, and is a prerequisite for hepatectomy planning and safe resection-margin determination. Existing deep learning approaches — 3D U-Net, nnU-Net, Swin UNETR, and Transformer variants such as LiverFormer — rely on intensity and spatial cues. Because adjacent segments exhibit nearly identical parenchymal intensity and boundaries are anatomically defined by vessels, these methods degrade precisely where precision matters most: near vascular structures. Prior vascular-aware techniques (Frangi vesselness filters, topology-preserving segmentation, vascular graph representations) have not been integrated into Couinaud frameworks, and standard Dice/Focal losses optimize voxel-wise accuracy without anatomical priors. VasGuideNet's premise is that Couinaud partitioning is fundamentally a vascularly governed functional partitioning, so the vessel tree should directly condition the segmentation.

A key scoping decision deserves emphasis: the method does not predict vessels. It requires vessel masks at test time, which is a meaningful deployment constraint discussed further below.

Architecture and topology encoding

The vascular encoding pipeline proceeds in three stages. First, the binary vessel mask is skeletonized, and $N$ keypoints are uniformly sampled along the centerline. Second, the Euclidean distance transform (EDT) of the vessel mask supplies a local radius $r_i \approx D(p_i)$ and a normalized orientation vector from the EDT gradient at each keypoint. Third, a k-nearest-neighbor graph over keypoints yields the adjacency structure, with initial node features formed by an MLP over position, radius, and orientation.

A stack of $L$ graph convolutional layers with normalized, self-looped adjacency produces node embeddings that capture vessel-tree connectivity and branching patterns. These embeddings serve as Key and Value tokens in a Transformer-style cross-attention, where the Stage-1 encoder output of the 3D-HCFormer backbone (Stem, four encoder stages, Context Bridge, Decoder) provides the Query tokens. Fused topology-aware features then propagate through the full encoder–decoder path.

Structural Contrastive Loss

SCL operates on high-confidence voxels only: anchors are selected above the 95th percentile of predicted confidence within each ground-truth class. Each anchor is contrasted, via a temperature-scaled cosine-similarity InfoNCE objective, against three negative sources: feature centers of other segments, the vessel center, and historical negatives drawn from a global memory bank. The inclusion of the vessel center as an explicit negative is the structurally distinctive element, pushing segment features away from vascular voxels where inter-segment boundaries are ambiguous.

The memory bank update strategy matters measurably. A FIFO queue ensures temporal diversity but discards high-quality entries; the proposed CATS (Confidence-Aware Time-Slice) strategy preferentially retains high-confidence items. CATS improves over FIFO by +0.59 Dice, +0.91 mIoU, and −0.30 RVD, supporting the claim that negative quality, not merely negative quantity, drives contrastive benefit in this setting.

Experimental results

Experiments use 5-fold cross-validation on Task08_HepaticVessel (443 CT scans, Couinaud labels following prior GLC-UNet annotations) and the private LASSD dataset (147 CT scans with expert-annotated Couinaud labels from a collaborating hospital). Metrics are DSC, mIoU, and RVD.

Method	Task08 DSC (%)	Task08 RVD (%)	LASSD DSC (%)	LASSD RVD (%)
G-UNETR++	59.32 ± 2.26	34.11 ± 3.89	65.14 ± 1.55	30.49 ± 8.51
UNETR	81.31 ± 1.60	27.50 ± 37.38	70.03 ± 6.11	33.04 ± 23.21
LiverFormer	71.51 ± 1.87	49.24 ± 31.83	57.23 ± 4.51	63.58 ± 12.30
SegMamba	79.05 ± 3.34	29.34 ± 15.34	69.00 ± 6.28	46.39 ± 27.09
SegFormer	79.51 ± 1.76	4.80 ± 0.85	73.09 ± 3.53	11.44 ± 2.12
3D-HCFormer	82.23 ± 0.87	3.50 ± 1.56	75.00 ± 4.01	11.28 ± 4.79
VasGuideNet	83.68 ± 0.82	1.68 ± 0.29	76.65 ± 3.33	7.08 ± 1.55

The most notable gains are volumetric rather than merely overlap-based: RVD drops from 3.50 to 1.68 on Task08 and from 11.28 to 7.08 on LASSD relative to 3D-HCFormer, with substantially tighter variance. The low standard deviation of VasGuideNet's RVD (0.29 on Task08) contrasts sharply with the high variance of UNETR (±37.38) and SegMamba (±15.34), indicating that topology guidance stabilizes volume calibration across anatomical variability — the specific failure mode the method targets.

Ablations isolate both components. Against the no-injection baseline (which is 3D-HCFormer), GCN–cross-attention fusion contributes +1.45 Dice, +2.06 mIoU, and −1.82 RVD; channel concatenation of topology features recovers much of the Dice gain (83.23) but less of the mIoU gain, while using the EDT map as an additive attention bias yields negligible benefit (82.25 Dice), showing that the relational structure captured by the GCN, not raw distance information, carries the signal. SCL contributes the largest single ablation effect: +2.79 Dice and +4.00 mIoU over the no-SCL baseline, exceeding the contribution of the fusion module alone.

Limitations and open questions

The central limitation is the test-time dependence on expert-annotated vessel masks. The paper explicitly frames its contribution as vascular guidance rather than vessel prediction, but this makes the reported gains conditional on annotation availability that is unrealistic at scale; the natural question — whether predicted or noisy vessel masks preserve the benefit — is not addressed. Second, evaluation on the private LASSD cohort (147 scans, single center) limits external validation of the generalization claims, and the Task08 Couinaud labels derive from a prior automatic annotation rather than independent expert review. Third, the ablation of fusion strategies and SCL update rules is conducted only on Task08; whether the CATS-over-FIFO margin persists on smaller datasets is untested. Finally, SCL hyperparameters (anchor percentile, temperature) are not ablated, leaving sensitivity uncharacterized.

Conclusion

VasGuideNet demonstrates that explicit GCN-encoded vascular topology, injected via cross-attention and reinforced by a vessel-aware contrastive objective, yields consistent improvements in Dice, mIoU, and — most distinctly — volume calibration and its variance for Couinaud segmentation. The strongest evidence is the ablation decomposition showing SCL and topology fusion as complementary, individually measurable contributions. The method's clinical applicability, however, remains tied to the availability of vessel annotations at inference, which is the principal open question for extending this approach beyond settings with expert vessel labeling.