MARVEL: Universal Murray's Law-informed Vessel Tree Segmentation and Topology Estimation

Yi Zhou, Thiara Sana Ahmed, Jacqueline Chua, Meng Wang, Qinrong Zhang, Alejandro F. Frangi, Huazhu Fu, Jun Cheng, Leopold Schmetterer, Bingyao TanarXiv 2026 · arXiv

TL;DR

MARVEL is a backbone-agnostic framework that integrates biophysical principles—specifically a universal formulation of Murray’s Law—into deep learning for vascular segmentation. By enforcing energy-optimal branching constraints during training, MARVEL produces physiologically plausible vessel trees that significantly improve the accuracy of downstream clinical tasks, such as diagnosing hypertension through hemodynamic simulation.

What Problem Does This Paper Address?

Accurate segmentation of vascular structures (like those in the eye, heart, or brain) is vital for medical diagnosis. However, contemporary deep learning models often treat segmentation as a simple "texture recognition" task. While they are good at identifying pixels, they frequently generate results that are biologically impossible: vessels might abruptly change thickness, disconnect randomly, or branch in ways that violate fluid dynamics.

These "topological errors" make the resulting maps useless for critical downstream applications like blood flow simulation or quantifying disease progression. The authors identify a fundamental disconnect between data-driven segmentation and the biophysical laws that govern how our bodies actually build vascular networks.

Methods

The core of the paper is the integration of Murray’s Law as a differentiable loss function. Murray’s Law states that for a vascular system to be energy-efficient, the radii of a parent vessel ($r_0$) and its daughter branches ($r_i$) must follow a specific power-law relationship: $r_0^\alpha = \sum r_i^\alpha$.

Key innovations include:

  • Universal Murray’s Law: Instead of assuming a fixed exponent (usually $\alpha=3$), the authors recognize that $\alpha$ varies based on vessel diameter due to effects like the Fahraeus-Lindqvist effect (where blood viscosity changes in narrow tubes). They created an empirical mapping between vessel width and the optimal exponent.
  • Differentiable Training Pipeline: To make these physical laws "readable" by a neural network, they implemented:
  • Soft Skeletonization: A method to extract vessel centerlines while maintaining gradient flow for backpropagation.
  • Radius Prediction: The network explicitly predicts a radius map alongside the segmentation mask.
  • Adaptive Murray Loss: A physics-informed loss function that penalizes branching patterns that deviate from the empirical width-exponent mapping.

![Fig.

  1. Illustration of Murray’s law across various vascular modalities. The Murray exponent α depends on the vessel radii ri.](/figures/art_de27d7d230b34aa795fe9f6210f48585)

Datasets

The authors evaluated MARVEL on eight public datasets across three distinct imaging modalities to prove its versatility:

  • 2D Retinal Fundus: RITE, LES-AV, F-AVSeg, and HRF-AV.
  • 3D Coronary CTA: ImageCAS and CCA.
  • 3D Cerebrovascular TOF-MRA: MIDAS and COSTA.

Key Insights and Ideas

  • Physics as a Regularizer: The biophysical prior acts as a powerful guide, especially in noisy or low-contrast regions where visual data is ambiguous. The model "knows" how a vessel should look based on its neighbors, even if the pixels are blurry.
  • Topology Matters for Clinical Utility: The paper demonstrates that "better pixels" (standard accuracy) don't always mean "better medicine." By preserving topological connectivity, MARVEL allows for more accurate hemodynamic simulations.
  • Robustness to Noise: Because the model is grounded in physical laws, it is significantly more resilient to image corruptions like blur or low brightness compared to purely data-driven baselines.

Key Results

MARVEL consistently outperformed state-of-the-art methods across all eight datasets in both segmentation accuracy (Dice score) and topological fidelity (clDice, Betti numbers).

  • Segmentation Gains: On the F-AVSeg dataset, MARVEL improved the Dice score by 2.77% over the best baseline.
  • Clinical Impact: In a hypertension classification task, a model using MARVEL-derived vessel maps achieved an AUC of 0.80, significantly outperforming the baseline AUC of 0.71 ($p < 0.001$). This was achieved by calculating the pressure difference between arteries and veins ($\Delta P_{AV}$), a biomarker that requires perfect vessel connectivity to compute.
  • Qualitative Improvement: As seen in Figure 3, MARVEL eliminates the "fragmented" look of traditional segmentation, resolving thin capillaries that other models miss or break.
Fig. 3. Qualitative comparison of retinal arterial and venule tree extraction methods across four datasets: (a) RITE, (b) LES-AV, (c) F-AVSeg, (d) HRF-AV.

Limitations and Future Work

The authors note that while Murray’s Law describes healthy energy-optimal transport, certain pathologies (like neovascularization in diabetes or intraretinal microvascular abnormalities) inherently violate these principles. Future iterations of MARVEL could include adaptive mechanisms to "relax" these physical constraints in regions identified as potentially pathological. Additionally, the authors aim to scale this hemodynamic validation from 2D retinal models to full 3D volumetric simulations that include entire capillary beds.

Summary updated

Abstract

Vascular circulation follows fundamental biophysical principles that optimize mass transport and metabolic energy expenditure, which can be effectively modeled by Murray's law. However, contemporary deep learning methods for vascular segmentation often neglect these biophysical constraints. This leads to physiologically implausible branching and misclassification vascular trees, rendering. These automated segmentation results are unreliable unreliable for downstream clinical tasks such as blood flow simulation or disease quantification. In this paper, we introduce MARVEL (Universal MurrAy's law-infoRmed Vessel sEgmentation and topoLogy estimation), a backbone-agnostic framework that integrates biophysical priors into vascular tree extraction. MARVEL combines per-pixel supervision with explicit radius predictions to enforce local bifurcation constraints derived from an empirical width-exponent mapping. We implement these constraints as differentiable regularizers during training to guide models toward physiologically consistent reconstructions. We evaluate MARVEL on eight public datasets across multiple vascular modalities and segmentation backbones. Results demonstrate MARVEL's superior performance in segmentation accuracy, topological consistency, and physiological plausibility. By converting segmented masks into graph-based hemodynamic simulations, we demonstrate that MARVEL preserves the subtle pathological narrowing and topological connectivity required to distinguish hypertensive from normotensive eyes. Results show that MARVEL significantly improves the classification of hypertension via arteriovenous pressure differences in the eye (p < 0.001), outperforming baseline models in both topological consistency and clinical predictive value.
BibTeX
@misc{zhou2026marvel,
  author = {Yi Zhou and Thiara Sana Ahmed and Jacqueline Chua and Meng Wang and Qinrong Zhang and Alejandro F. Frangi and Huazhu Fu and Jun Cheng and Leopold Schmetterer and Bingyao Tan},
  title = {{MARVEL: Universal Murray's Law-informed Vessel Tree Segmentation and Topology Estimation}},
  year = {2026},
  eprint = {2605.25363},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2605.25363},
}
Figure