NeuroVascU-Net: T1CE MRI Vessel Segmentation

• The paper introduces NeuroVascU-Net, a unified deep learning approach that leverages a dilated U-Net backbone with advanced multi-scale and cross-domain adaptive feature fusion for precise 3D vessel segmentation.
• It integrates specialized MSC²F and CDA²F modules that fuse multi-scale, frequency, and structural features, improving both segmentation accuracy and computational efficiency.
• Extensive ablation studies and quantitative evaluations demonstrate high Dice scores and improved precision, offering a clinically feasible solution for neuro-oncology vascular mapping.

NeuroVascU-Net is a unified deep learning architecture tailored for precise three-dimensional segmentation of cerebral vasculature in T1-weighted contrast-enhanced magnetic resonance imaging (T1CE MRI), specifically within neuro-oncology patient populations. Addressing a longstanding focus on time-of-flight magnetic resonance angiography (TOF-MRA) data in prior vessel segmentation literature, NeuroVascU-Net delivers high-fidelity segmentation exclusively from clinically prevalent T1CE MRI volumes. It accomplishes this by incorporating advanced multi-scale and cross-domain adaptive feature fusion modules into a fundamentally dilated U-Net backbone, balancing segmentation accuracy, computational efficiency, and clinical practicality (Vayeghan et al., 23 Nov 2025).

1. Architectural Overview

NeuroVascU-Net is built on a five-level, three-dimensional U-Net backbone (input: $192\times192\times128$ voxels, single-channel). The encoder comprises consecutive pairs of $3\times3\times3$ dilated convolutions (default dilation $(1,1,1)$), each followed by Batch Normalization and ReLU activation, and down-sampling via $2\times2\times2$ max-pooling. The encoder’s channel widths progress as $C=\{16,32,64,128,256\}$. Specialized modules replace standard convolutions at deeper levels: the Cross-Domain Adaptive Feature Fusion ($\mathrm{CDA}^2\mathrm{F}$) module appears at level 4, and the Multi-Scale Contextual Feature Fusion ($\mathrm{MSC}^2\mathrm{F}$) module is inserted at the bottleneck (level 5).

The decoder symmetrically mirrors the encoder architecture, using $2\times2\times2$ transposed convolutions for up-sampling and attention-gated grid skip connections. $\mathrm{CDA}^2\mathrm{F}$ is re-applied at decoder level 4. The final $1\times1\times1$ convolution maps the feature space to dense per-voxel vessel probabilities. NeuroVascU-Net contains 12.43 million trainable parameters.

2. Advanced Feature Fusion Modules

2.1 Multi-Scale Contextual Feature Fusion (MSC²F)

The $\mathrm{MSC}^2\mathrm{F}$ module at the bottleneck fuses multi-scale, frequency, and structural information. Starting with atrous spatial pyramid pooling (ASPP) using three anisotropic dilation rates $d\in\{(1,1,1),(1,2,2),(1,3,3)\}$, the outputs are concatenated across channels. An edge token is computed using a 3D Laplacian-of-Gaussian kernel and a frequency token is derived by applying a learnable spectral mask in the Fourier domain. The bottleneck feature, its edge and frequency cues, and the skip connection are concatenated and refined using a depthwise 3D convolution and Efficient Channel Attention (ECA), followed by a $1\times1\times1$ convolution and residual summation for final output:

$$F_{\mathrm{MSC}^2\mathrm{F}} = \mathrm{Conv}_{1\times1\times1}(F_{\mathrm{ASPP}}) + \mathrm{Conv}_{1\times1\times1}(\widehat{Y})$$

2.2 Cross-Domain Adaptive Feature Fusion (CDA²F)

The $\mathrm{CDA}^2\mathrm{F}$ module (in both encoder and decoder level 4) fuses domain-specific feature representations using four parallel processing branches: 3D involution (location-specific spatial filtering), frequency-spatial attention, spherical CNN (rotationally equivariant features), and a depthwise ConvNeXt block (kernel $5\times5\times5$, GELU activation). Three branches are summed, rescaled, and merged with the ConvNeXt output, followed by stochastic depth with residual connections and a gated axial transformer that applies self-attention sequentially along each volume axis. Output gating uses a learned sigmoid mask.

These modules ensure hierarchical multi-scale integration and cross-domain robustness, supporting accurate extraction of both large vessels and fine distal structures.

3. Training Data, Preprocessing, and Optimization

The model was trained on a curated dataset of 137 patients scanned at Rasoul Akram Hospital, each with $\sim$160 axial T1CE MRI slices ($256\times256$, 1 mm in-plane). Manual 3D vessel annotations were produced using 3D Slicer by a board-certified functional neurosurgeon. Preprocessing steps included skull-stripping (HD-BET), N4 bias-field correction (SimpleITK), intensity normalization, and resizing to $192\times192\times128$ voxels. Data augmentation comprised random y-axis flipping (30% probability) and background Gaussian noise ($\mu=0$, $\sigma=0.01$).

The hybrid training loss was:

$$\mathcal{L}_\mathrm{Total} = 2.0\,\mathcal{L}_\mathrm{WCE} + 1.0\,\mathcal{L}_\mathrm{Dice}$$

with weighted cross-entropy (WCE, vessel:background weight ratio $8.546:1$) and standard Dice loss, as defined in the manuscript.

Optimization used the Adam optimizer (learning rate $8{\times}10^{-5}$, dropout $0.2$, batch size $2$) with early stopping on validation loss, and sliding-window 3D inference. The data were split into 100 training, 10 validation, and 27 test cases.

4. Quantitative and Qualitative Performance

On the held-out test set, NeuroVascU-Net achieved:

Metric	Value
Dice (DSC)	0.8609
Jaccard (JI)	0.7582
Sensitivity	0.8456
Specificity	0.9982
Precision	0.8841

Comparison with Swin U-NetR (15.7M parameters) showed equivalent Dice (0.8600) but notably higher precision for NeuroVascU-Net (0.8841 vs 0.8454), despite using fewer parameters (12.4M). Qualitative analysis demonstrated accurate delineation of both major vessels and fine arterioles/venules, with minimal false positives outside white matter and limited false negatives at vessel boundaries. Three-dimensional overlays confirmed preservation of vascular connectivity in tortuous and small-diameter segments.

Computational efficiency is maintained, with average inference time of 3840 s per volume (sliding-window) on a single Tesla P100, and GPU memory consumption compatible with standard 16 GB cards at a batch size of 2.

5. Ablation Studies and Component Analysis

Ablation experiments provided evidence for the necessity of the MSC²F and CDA²F modules. Replacing MSC²F with plain dilated convolutions decreased Dice by $\sim$1.9% (from 0.8609 to $\sim$0.842) and reduced capillary-level segmentation sensitivity. Removing CDA²F or reverting to standard dilated-convolution blocks at level 4 resulted in reductions of approximately 1.2% in Dice and 2.0% in precision. Loss function analysis established the superiority of the hybrid WCE+Dice approach; focal loss did not yield comparable performance. The architecture displayed sensitivity to spatial detail: a reduction in input resolution to $160\times160\times128$ voxels led to a $\sim$0.02 decrease in Dice, indicating reliance on high-fidelity input for accurate fine-vessel detection.

6. Significance and Comparative Perspective

NeuroVascU-Net is the first architecture explicitly developed for vessel segmentation in conventional T1CE MRI in neuro-oncology, addressing a gap in a field where most automated methods employ TOF-MRA. Its reach extends to neurosurgical planning, offering fast and accurate vessel delineation without the need for time-consuming manual annotation or specialized imaging protocols.

The combination of a dilated U-Net base with advanced multi-scale and cross-domain feature fusion yields competitive or superior performance compared to transformer-based methods, while reducing parameter count and maintaining practical computational requirements (Vayeghan et al., 23 Nov 2025). This positions NeuroVascU-Net as a clinically feasible solution for computer-assisted preoperative vascular mapping.