Technical Summary: RadicularSEUNet for SDAVF Localization

Problem Statement

Spinal dural arteriovenous fistula (SDAVF) is the most common spinal vascular malformation, yet its diagnosis remains challenging. The current gold standard, spinal digital subtraction angiography (DSA), is invasive, time-consuming, exposes patients to radiation, and suffers from a high rate of false negatives due to technical difficulties in catheterizing tortuous segmental arteries, particularly in elderly patients. While contrast-enhanced MR angiography (CEMRA) offers a noninvasive alternative, it lacks the spatial resolution to directly visualize the fistulous zone (<0.5 mm). Consequently, CEMRA often fails to demonstrate the complete angioarchitecture, making it difficult for clinicians to localize the lesion based on arterial flow direction as they would with DSA. Furthermore, existing CEMRA localization methods rely on threshold-based postprocessing, which struggles to distinguish small radicular veins from epidural venous plexuses or noise, leading to incomplete visualization and missed diagnoses.

Methodology

Data Strategy and Synthetic Generation

Recognizing the scarcity of real-world SDAVF cases and the difficulty of manual annotation, the authors developed a synthetic SDAVF generation pipeline to create a large-scale training dataset (n=2,000). This pipeline utilizes anatomical constraints derived from spinal cord surface, dura mater, and intervertebral foraminal anatomy:

• Anatomical Scaffolding: Masks of bone, discs, and the spinal canal are segmented from pre-contrast MRI. The spinal canal mask serves as a proxy for the dural sac.
• Vein Synthesis: A dorsal longitudinal trunk is modeled with stochastic morphological modes (serpentine, coiled) and secondary branches (loops, hairpins). A single radicular vein is synthesized to connect the lateral dura to the surface venous network, adhering to specific anatomical rules (e.g., connection within 3 disc levels, specific angular sectors).
• Intensity Modeling: Voxel intensities are modulated based on reference statistics from real SDAVF cases, incorporating axial variation and radial attenuation to mimic realistic vessel contrast and avoid homogeneous cross-sections.

Model Architecture: RadicularSEUNet

The study introduces RadicularSEUNet, a deep learning model based on a 3D Squeeze-and-Excitation U-Net (SE-U-Net) framework designed for the 3D segmentation of the subdural draining vein.

• Input: A two-channel input comprising the original 3D CEMRA image and a corresponding dura mask (derived from the spinal canal).
• Architecture: The network features four downsampling and four upsampling stages. Encoder blocks utilize residual 3D convolutions with instance normalization, leaky ReLU activation, and Squeeze-and-Excitation (SE) modules to recalibrate channel-wise features. The decoder employs transposed convolutions and skip connections.
• Loss Function: Training utilizes a hybrid loss combining focal Tversky loss (to address class imbalance and emphasize difficult vessel voxels) and soft clDice loss (to promote topological continuity in tubular structures).
• Training: The model was trained on 1,600 synthetic cases, validated on 200, and tested on 200 synthetic cases.

Clinical Workflow

For clinical application, the workflow involves:

1. Classification: Generating a dura-restricted 2D Maximum Intensity Projection (MIP) to differentiate SDAVF from other malformations (e.g., AVMs, pial AVFs) based on the presence or absence of a hairpin-shaped feeding artery.
2. Localization: If SDAVF is suspected, RadicularSEUNet segments the 3D intradural venous network. This mask is fused with vertebral and dura masks to visualize the junction point between the radicular vein and the dura mater.
3. Level Determination: A blinded physician identifies the segmental level based on the vertebral pedicle projection relative to the dural connection point.

Key Results

Synthetic Data Performance

On the independent synthetic test set (n=200), RadicularSEUNet demonstrated high segmentation fidelity:

• Median Dice Coefficient: 0.986
• Median Centerline Dice (clDice): 0.986
• Median F0.5 Score: 0.992
• Composite Score: 0.988
  The model showed stable performance across various thresholds, indicating robustness in segmenting the complex intradural venous network.

Clinical Cohort Performance

The model was evaluated on a retrospective cohort of 31 real CEMRA cases (21 SDAVF, 6 pial AVFs, 2 SAVMs, 2 DSA-negative).

• Differentiation: The combined workflow (MIP + RadicularSEUNet) correctly differentiated SDAVF from other vascular malformation types.
• Localization Accuracy: The model achieved 100% accuracy in level localization for SDAVF cases.
• Challenging Cases:
  • Upper/Mid-Thoracic: In 7 cases with short, thin radicular veins (typically difficult to segment), the model correctly identified the dural connection level in 100% of cases (7/7), whereas 2D MIP images failed to determine the location in 3 cases (42.9%).
  • DSA-Negative Cases: In two DSA-negative cases (diagnosed as occult tethered spinal cord), RadicularSEUNet identified typical vein patterns but correctly distinguished them from SDAVF based on low voxel counts (sparse vesselness), preventing false positives.

Significance and Claims

The paper claims that RadicularSEUNet enables a novel, vein-based, noninvasive, and anatomically intuitive strategy for SDAVF localization. By shifting the focus from the elusive arterial feeder to the more visible and larger-diameter radicular vein, the method overcomes the spatial resolution limitations of CEMRA.

Key contributions highlighted by the authors include:

1. Overcoming Resolution Limits: The model successfully segments the radicular vein even when the fistulous zone is not directly visible, providing a reliable surrogate for localization.
2. Automation and Objectivity: Unlike threshold-based segmentation which is operator-dependent, RadicularSEUNet offers a fully automated solution that reduces inter-observer variability.
3. Clinical Utility: The approach achieved 100% localization accuracy in a real-world cohort, suggesting it can reduce missed diagnoses, lower radiation exposure by minimizing the need for repeated or extensive DSA, and potentially serve as a standalone diagnostic tool.
4. Anatomical Visualization: By reconstructing the connection between the radicular vein and the dura mater alongside vertebral anatomy, the method provides a visualization that aligns with surgical targets (interruption of the draining vein) and physician workflows.

The authors acknowledge limitations, including the need for validation in larger prospective cohorts, the use of the spinal canal mask as a dura proxy, and occasional discontinuity in very long, slender veins. However, they posit that this work represents a significant step toward a radiation-free diagnostic and therapeutic pathway for SDAVF.