Stroke Locus Net: Vascular Occlusion Localization

• Stroke Locus Net is a deep learning framework that combines lesion segmentation and synthetic angiography to precisely localize occluded vessels in stroke patients.
• It employs a two-branch architecture integrating nnU-Net for lesion detection and pGAN for synthesizing MRA images, enhancing overall vascular mapping accuracy.
• The system leverages atlas-based arterial territory mapping and evaluates performance with metrics like DSC, PSNR, and SSIM, supporting rapid, automated stroke diagnosis.

Stroke Locus Net encompasses a class of deep learning architectures and computational pipelines aimed at localizing the vascular origin of ischemic stroke—specifically, mapping the site of vessel occlusion—using MRI modalities. Unlike earlier work that focused solely on lesion segmentation, Stroke Locus Net integrates end-to-end lesion detection, vascular territory mapping, and virtual angiography, resulting in comprehensive occluded vessel localization from MRI data without direct angiographic input (Hamad et al., 11 Oct 2025).

1. System Architecture and Branch Design

Stroke Locus Net is structured as a two-branch deep learning pipeline:

• Segmentation Branch: Utilizes nnU-Net for ischemic lesion detection on MRI scans. nnU-Net self-configures preprocessing, architecture, and training parameters to output a binary mask of infarcted tissue, typically from T1- or T2-weighted images.
• Generation Branch: Applies a pixel-wise conditional Generative Adversarial Network (pGAN) to synthesize MRA (Magnetic Resonance Angiography) images from standard MRI scans. This enables vessel visualization and segmentation even in the absence of a dedicated MRA scan.

Integration between the branches provides both anatomical (lesion) and vascular (vessel) delineations, facilitating the localization of the occluded vessel responsible for the stroke.

2. Lesion Segmentation and Arterial Territory Mapping

Lesion segmentation is accomplished using nnU-Net, trained on annotated stroke MRI datasets (e.g., ATLAS). The binary lesion mask is then spatially registered to a probabilistic arterial atlas using FSL’s FLIRT for rigid alignment. The atlas divides the brain vasculature into ten arterial territories (e.g., right and left middle cerebral artery, anterior and posterior cerebral artery domains, anterior choroidal artery, etc.).

Stroke Locus Net computes the percentage and absolute voxel overlap of the segmented lesion with each arterial territory, assigning the stroke’s vascular locus to the region with maximal overlap. This inferential step directly addresses the clinical need to localize the site of arterial occlusion, enabling targeted therapy planning.

3. Synthetic Angiography and Vessel Segmentation

The generation branch addresses the challenge of absent MRA data by using a pixel-wise conditional GAN (pGAN) to transform structural MRI contrast (e.g., T2) into synthetic MRA images. The GAN setup includes:

• Generator ($G$): Learns to convert input MRI slices ($x_k$) to synthetic MRA predictions ( $G(x_k)$ ).
• Discriminator ($D$): Distinguishes real MRA from synthetic output.
• Objective/Loss:

$$L_{\text{pGAN}} = L_{\text{condGAN}_k}(D, G) + \lambda\,L_{L1_k}(G) + \lambda_{\text{perc}}\,L_{\text{perc}_k}(G)$$

where: - $L_{\text{condGAN}_k}(D,G)$ is least-squares adversarial loss, - $L_{L1_k}(G)$ is pixel-wise $L_1$ loss, - $L_{\text{perc}_k}(G)$ is perceptual loss using VGG16 features, - $\lambda,\,\lambda_{\text{perc}}$ are weighting terms.

Vessel segmentation is performed on the generated MRA images using an established four-class segmentation network. Fine-grained arterial regions from the atlas are merged into broader vessel categories to match the output classes. This process, when paired with the lesion mask and vascular territory information, highlights the most probable site of vessel occlusion.

4. Performance Metrics and Evaluation

Key empirical results include:

• Vessel Localization: MRA-based vessel segmentation achieved a Dice Similarity Coefficient (DSC) of ~0.84.
• MRI-based Vessel Segmentation: Direct segmentation from MRI achieved a DSC of ~0.75, indicating the added value of synthetic MRA for vessel mapping.
• Image Synthesis: For T2-to-MRA translation via pGAN, Peak Signal-to-Noise Ratio (PSNR) was 17.82±2.26 and Structural Similarity Index Measure (SSIM) was 0.31±0.01. These scores are lower than for within-MRI contrast translation, reflecting the intrinsic difficulty of recovering vascular detail from T2 anatomy alone.

The fusion of MRI and synthetic MRA images further enhanced vascular structure visualization, improving vessel localization accuracy and interpretability.

5. Clinical and Methodological Implications

Stroke Locus Net’s integration of lesion segmentation, arterial territory assignment, and vessel synthesis has the following clinical implications:

• Accelerated Diagnosis: Enables occluded vessel localization using only standard MRI, facilitating faster triage and targeted management particularly where CTA/MRA is unavailable or slow to acquire.
• Comprehensive Assessment: Combines tissue damage mapping and vascular assessment in a single pipeline, supporting detailed stroke characterization.
• Reduced Imaging Burden: Mitigates the need for high-resolution vascular imaging (CT angiogram or contrast-enhanced MRA), which is especially valuable in emergency or resource-limited settings.
• Automated Arterial Assignment: The pipeline automatically assigns lesions to specific arterial territories via atlas overlap, supporting consistent and reproducible diagnosis.

6. Limitations and Future Directions

Current limitations exist primarily in the generation branch:

• Synthesized MRA can capture overall spatial and topological information, but produces lower structural fidelity than conventional MRA, limiting the direct use of synthetic data for fine vessel segmentation.
• The system employs four-class vessel segmentations due to the limitations of the underlying vessel segmentation model, necessitating pooling of arterial atlas regions.

Future improvements may include:

• Enhanced GAN architectures or hybrid approaches to improve extraction of fine vascular details.
• Integration of more granular or context-based mapping from lesions to sub-territorial vasculature.
• Extension to additional MRI contrasts or incorporation of multi-modal imaging.

7. Algorithms and Formalisms

The vessel localization process in Stroke Locus Net involves an explicit overlap mapping:

• For each arterial territory $A_i$, compute

$O(A_i) = \frac{|A_i \cap L|}{|L|}$

where $L$ is the lesion mask.

• Assign the locus to the $A_i$ with maximal $O(A_i)$.
• Vessel segmentation merges atlas regions to four classes matching the existing vessel segmentation model output (e.g., MCAL+MCAR become “MCA”).

The overall workflow fuses segmentation, registration, GAN-based synthesis, vessel segmentation, and probabilistic assignment in a unified computational framework.

---

In sum, Stroke Locus Net operationalizes the direct, automated localization of occluded vessels from MRI by integrating multi-branch deep learning, probabilistic vascular atlases, and synthetic angiography, thereby advancing the precision and timeliness of ischemic stroke assessment (Hamad et al., 11 Oct 2025).