MVis-Fold: A Three-Dimensional Microvascular Structure Inference Model for Super-Resolution Ultrasound

Abstract: Super-resolution ultrasound (SRUS) technology has overcome the resolution limitations of conventional ultrasound, enabling micrometer-scale imaging of microvasculature. However, due to the nature of imaging principles, three-dimensional reconstruction of microvasculature from SRUS remains an open challenge. We developed microvascular visualization fold (MVis-Fold), an innovative three-dimensional microvascular reconstruction model that integrates a cross-scale network architecture. This model can perform high-fidelity inference and reconstruction of three-dimensional microvascular networks from two-dimensional SRUS images. It precisely calculates key parameters in three-dimensional space that traditional two-dimensional SRUS cannot readily obtain. We validated the model's accuracy and reliability in three-dimensional microvascular reconstruction of solid tumors. This study establishes a foundation for three-dimensional quantitative analysis of microvasculature. It provides new tools and methods for diagnosis and monitoring of various diseases.

• The paper introduces MVis-Fold, a deep learning-based framework that leverages Transformer and graph neural network architectures to enable 3D microvascular reconstruction from 2D SRUS images.
• It fuses multi-modal 2D inputs to quantify complex vascular features, reducing vessel density and diameter estimation errors by up to 1,353-fold and 55-fold respectively.
• Its robust performance in murine models and rapid inference time underscore its potential for noninvasive, dynamic imaging applications in oncology and beyond.

MVis-Fold: A Three-Dimensional Microvascular Inference Model for Super-Resolution Ultrasound

Introduction

The MVis-Fold model addresses the critical limitation of two-dimensional (2D) super-resolution ultrasound (SRUS) microvascular imaging by providing a robust three-dimensional (3D) reconstruction framework for microvascular networks. Conventional imaging modalities such as CT and MRI are constrained by spatial resolution insufficient for microvasculature, and SRUS, while achieving micrometer-scale detail, is fundamentally restricted to the 2D plane. This dimensional constraint impedes accurate quantification of pivotal microvascular features and hinders dynamic, in vivo assessment crucial for oncology, neurology, and systemic disease applications.

Technical Approach

MVis-Fold leverages deep learning and cross-scale Transformer-based architectures, directly inspired by AlphaFold3’s success in 3D protein structure inference. The model incorporates multi-head self-attention mechanisms capable of integrating SRUS-derived flow direction, flow angle, and morphological data, thus capturing implicit spatial dependencies not accessible through 2D observation alone. The core architecture employs:

• Hierarchical feature extraction via self-attention modules across modalities (grayscale, flow maps, etc.).
• Multi-scale feature pyramids enabling fusion and representation of vessels of variable calibers.
• A 3D topological rationalization module with graph neural networks to enforce vascular connectivity priors.
• Bayesian uncertainty estimation for spatiotemporal consistency.

The input is a set of 2D SRUS images and their derivatives, outputting high-fidelity 3D microvascular reconstructions with comprehensive uncertainty maps.

Validation and Numerical Performance

MVis-Fold was extensively validated in a well-powered murine solid tumor model, utilizing over 16,000 SRUS-derived datasets with pathological gold standard comparison. The model demonstrated:

• Dice coefficient: $0.959 \pm 0.034$
• Sensitivity: $0.951 \pm 0.038$
• Specificity: $0.957 \pm 0.025$
• Morphological parameter Pearson correlation with histopathology: $r = 0.892$, $p<0.001$
• Inference speed: $8.3 \pm 0.4$ seconds per volume

These results represent an order-of-magnitude improvement over classical (SparseNeuS, OpenLRM, TripoSR) and deep learning baselines. Notably, vessel density and diameter estimation errors were reduced by 1,353-fold and 55-fold respectively compared to state-of-the-art 2D SRUS, achieving near-pathological precision. This degree of fidelity in quantifying parameters such as vessel density, mean diameter, and topological connectivity establishes MVis-Fold as a clear advance for in vivo microvascular science.

Implications for Quantitative Microvascular Imaging

MVis-Fold’s methodological innovation is significant in that it enables multidimensional microvascular analysis. Beyond basic 2D density estimation, it supports quantitative extraction of 3D topological (branching angle, network heterogeneity), morphological (surface, volume), and even hemodynamic metrics (perfusion index, blood flow resistance). Crucially, this is achieved in a non-invasive and dynamic fashion, contrasting sharply with the static, invasive, and post hoc nature of histopathological techniques.

The model’s superior generalization is supported by its multi-modal input fusion and topological consistency constraints, suggesting broad applicability to various disease settings and imaging contexts. The architecture sets the stage for future integration of additional imaging physics (e.g., photoacoustic or hybrid modalities), with transfer to human clinical data strongly implied albeit not yet definitively demonstrated.

Limitations and Future Prospects

The primary empirical limitation is the current restriction of validation to murine oncology models. Translation to human clinical scenarios will require tackling practical challenges such as robust acquisition of multiplane SRUS data, standardized workflow integration, and clinical interpretability of topological and hemodynamic outputs.

Potential extensions include application to neurovascular, ischemic, and degenerative models, adaptation for high-throughput quantitative biomarker discovery, and integration into real-time clinical decision support systems. Algorithmic improvements might exploit even deeper multimodal fusion, interpretable Transformer variants, or joint training with downstream diagnostic or prognostic endpoints.

Conclusion

MVis-Fold constitutes a comprehensive and technically rigorous solution for 3D reconstruction of microvascular networks from 2D SRUS data. Its high accuracy, generalizability, and computational efficiency provide crucial advances for noninvasive, dynamic microvascular imaging. This model transitions medical ultrasonography from a primarily qualitative, 2D paradigm to a quantitative, multidimensional analytic tool. Future developments will determine its ultimate translational potential in precision medicine and broader AI-driven medical imaging.