NeuroVascU-Net: A Unified Multi-Scale and Cross-Domain Adaptive Feature Fusion U-Net for Precise 3D Segmentation of Brain Vessels in Contrast-Enhanced T1 MRI

TL;DR

NeuroVascU-Net tackles the challenge of segmenting cerebral vessels directly from contrast-enhanced T1 MRI in brain tumor patients, a task previously dominated by TOF-MRA approaches. The authors introduce two specialized modules, MSC^2F at the bottleneck and CDA^2F at deeper levels, to fuse multi-scale, edge, frequency, and domain-specific features within a compact 3D U-Net framework. On a curated 137-patient T1CE dataset annotated by a neurosurgeon, the model achieves a Dice score of 0.8609 and precision of 0.8841 with 12.4M parameters, outperforming several transformer-based and CNN baselines. The work demonstrates that task-specific architectural biases can deliver high segmentation accuracy with practical computational efficiency, enabling integration into computer-assisted neurosurgical workflows, while noting the need for multi-center validation and possible extensions to sub-voxel vessels and intraoperative optimization.

Abstract

Precise 3D segmentation of cerebral vasculature from T1-weighted contrast-enhanced (T1CE) MRI is crucial for safe neurosurgical planning. Manual delineation is time-consuming and prone to inter-observer variability, while current automated methods often trade accuracy for computational cost, limiting clinical use. We present NeuroVascU-Net, the first deep learning architecture specifically designed to segment cerebrovascular structures directly from clinically standard T1CE MRI in neuro-oncology patients, addressing a gap in prior work dominated by TOF-MRA-based approaches. NeuroVascU-Net builds on a dilated U-Net and integrates two specialized modules: a Multi-Scale Contextual Feature Fusion ($MSC^2F$) module at the bottleneck and a Cross-Domain Adaptive Feature Fusion ($CDA^2F$) module at deeper hierarchical layers. $MSC^2F$ captures both local and global information via multi-scale dilated convolutions, while $CDA^2F$ dynamically integrates domain-specific features, enhancing representation while keeping computation low. The model was trained and validated on a curated dataset of T1CE scans from 137 brain tumor biopsy patients, annotated by a board-certified functional neurosurgeon. NeuroVascU-Net achieved a Dice score of 0.8609 and precision of 0.8841, accurately segmenting both major and fine vascular structures. Notably, it requires only 12.4M parameters, significantly fewer than transformer-based models such as Swin U-NetR. This balance of accuracy and efficiency positions NeuroVascU-Net as a practical solution for computer-assisted neurosurgical planning.

Figures (5)

• Figure 1: Comparison of vascular and tissue visualization in TOF-MRA and T1CE MRI. (a) TOF-MRA highlights large, fast-flowing arteries with limited soft tissue contrast and poor visibility of small or slow-flow vessels. (b) T1CE MRI provides enhanced anatomical detail and visualizes both arteries and veins, including smaller vessels near pathological tissue, but introduces complexity in vessel segmentation due to nonspecific enhancement.
• Figure 2: Segmentation process and proposed NeuroVascU-Net network structure.
• Figure 3: Detailed structure blocks for NeuroVascU-Net: (a) $MSC^2F$ block; (b) $CDA^2F$ block.
• Figure 4: NeuroVascU-Net performance over epochs (a) Train and overall validation loss. (b) Train and overall validation DSC, the lower training DSC in comparison to validation is caused by sliding window inference for validation data.
• Figure 5: Visualization of brain vessel segmentation performance for each DL model on the T1CE dataset. The first column shows the ground truth (GT), and the following columns display model predictions. Rows 1–3 show sample prediction vs. GT slices from validation data, while the fourth row presents a 3D overlay with opacity. True Positives (TP) in orange, False Positives (FP) in red, and False Negatives (FN) in green are highlighted in each image.