Intracranial Aneurysm Detection

• Intracranial Aneurysm Detection is the process of identifying abnormal cerebral artery dilations that pose high risks for rupture and hemorrhage.
• Recent advances combine deep learning, graph neural networks, and synthetic augmentation to enhance sensitivity and reduce false positives.
• Multimodal imaging and annotation strategies integrate volumetric and surface data to address challenges posed by anatomical variability and low contrast.

Intracranial aneurysm detection is the process of identifying and characterizing abnormal focal dilations of cerebral arteries, which carry significant risk due to potential rupture and resultant subarachnoid hemorrhage. In contemporary research and clinical practice, detection workflows must contend with anatomical variability, imaging noise, low lesion-to-background contrast, and substantial class imbalance. Recent advances leverage combinations of geometric priors, deep learning architectures, synthetic data augmentation, and cross-domain feature transfer to address these challenges and improve diagnostic sensitivity, specificity, and interpretability.

1. Data Representations and Annotation Strategies

Modern intracranial aneurysm detection employs diverse data representations to exploit salient geometric and anatomical variation:

• Surface Meshes and Point Clouds: Many frameworks leverage triangulated vascular surfaces extracted from angiographic scans (DSA, CTA, TOF-MRA) via algorithms such as marching cubes or multi-threshold reconstruction (Yang et al., 2020, Yang et al., 2020). Surface patches or local point clouds are formed by geodesic sampling (e.g., 3000 points per patch within a threshold along the mesh (Bizjak et al., 2020)), enabling both local and global context extraction.
• Volumetric Data: Volumetric patches (typically $64^3$ or $96^3$ voxels) provide full contextual information at high spatial resolution, suitable for 3D U-Net (Indrakanti et al., 30 Aug 2024), nnDetection/Retina U-Net (Orouskhani et al., 2023), and multi-task UNet variants (Rainville et al., 1 Aug 2025).
• Expert Annotations/Weak Labels: Ground truth is constructed from expert manual segmentations (possibly in the form of coarse spheres or surface "painting" for 3D models (Yang et al., 2020, Rainville et al., 1 Aug 2025)) and, for large datasets, may be complemented with automated procedures for consistent labeling.

Annotation complexity is compounded by the small size of aneurysms and non-manifold mesh artifacts inherent to clinical reconstructions. Synthetic datasets and expert-guided corrections (e.g., using MeshLab filters and shortest geodesic boundary marking (Yang et al., 2020, Nader et al., 4 Nov 2024)) play a crucial role in quality assurance and large-scale training.

2. Deep Learning Architectures for Detection and Segmentation

Aneurysm detection frameworks increasingly couple advanced neural architectures with geometric priors for robust performance:

• Graph Neural Networks (GNNs) and Surface-Based Pipelines: GraphNet (Ma et al., 2019) processes 3D triangulated surfaces using GCN blocks to extract local node-level and global pooled features for rupture classification and spatial segmentation:

$$H^{(l+1)} = \sigma(\tilde{D}^{-1/2} \tilde{A} \tilde{D}^{-1/2} H^{(l)} W)$$

where $\tilde{A}$ incorporates self-connections and $\tilde{D}$ is the degree matrix. Dual-branch outputs provide whole-aneurysm status and surface segmentation, with a reported mean DSC of $0.88$.

• Point Cloud Networks: PointNet/PointNet++ classifiers, often regularized for invariance to rotation, translation, and scale, discriminate aneurysm-containing patches with accuracy surpassing 98% on healthy segments and F1-scores above 0.90 given sufficient sampling density (Yang et al., 2020, Bizjak et al., 2020). Hybrid strategies combine local and global pooling for fine-grained segmentation (see PN++g/PointConv for enhanced boundary detection).
• Two-Step Surface and Fragment Approaches: Multistage workflows first classify small surface fragments or local patches for aneurysm presence (e.g., hierarchical PointNet++), and, upon positive detection, segment the aneurysm surface using models such as SO-Net. This pipeline achieved a Dice coefficient near 0.72, outperforming volumetric CNNs by over 25 percentage points (Yang et al., 2020).
• Volumetric and Multi-Task Networks: 3D U-Net and self-configuring nnU-Net structures are widely adopted (Indrakanti et al., 30 Aug 2024, Orouskhani et al., 2023). These encode multi-scale context and prioritize features relevant to the task. Multi-task UNet variants (Rainville et al., 1 Aug 2025) integrate vesselness priors using Frangi filters, driving attention to vascular regions in both encoding and decoding stages.
• False-Positive Reduction and Post-Processing: Models such as CADIA (Wu et al., 2020) use a 3D region proposal network (RPN) to scan the input volume followed by a DenseNet-based false positive reduction (FPR) module, producing lesion- or patient-level FROC curves. Further, automated anatomy-based post-processing harnesses artery-vein segmentation masks to eliminate false detections in non-arterial domains (reducing FP/case by over 50% in some setups (Kim et al., 1 Jul 2025)).

3. Synthetic Data Augmentation and Generative Modeling

The limited real-world annotated datasets for IA detection are frequently augmented by carefully constructed synthetic datasets:

• Synthetic Vascular Models: Full cerebral trees, arteries, bifurcations, and aneurysms are modeled by spline interpolation of centerlines, 3D convolution with deformed kernels, and insertion of elastically distorted spherical sacs for aneurysms (Nader et al., 27 Mar 2024, Nader et al., 4 Nov 2024). Gray-level and noise statistics of real TOF-MRA acquisitions are analytically replicated—e.g., via Gaussian filtering and Bienaymé's identity to match conditional variance: $$\sigma_f \approx \frac{\sigma_0}{2 \sigma_G \sqrt{\pi}}$$. This faithful modeling of vessel, bifurcation, aneurysm, and noise variability supports high-fidelity 3D CNN training.
• Effect of Augmentation: Experiments show that augmenting real patches with thousands of synthetic vascular patches increases detection sensitivity from 75.6% to nearly 89%, with primarily modest false positive penalties. Sensitivity for sub-2mm aneurysms benefits from synthetic cases, resulting in an improvement from 51% to 76% (Nader et al., 27 Mar 2024).
• Controllable Mesh Generation: The AneuG framework (Ding et al., 15 May 2025) employs a two-stage VAE driven by Graph Harmonic Deformation (GHD) tokens for aneurysm sac generation, with parent vessels synthesized in a correlated latent space. Clinical morphometrics (neck width, aspect ratio, lobulation index) and morphing energy alignment constrain synthetic meshes, permitting targeted studies of shape-dependent blood flow and rupture risk.
• Hemodynamics-Integrated Datasets: Aneumo provides paired 3D models, segmentation masks, and CFD-derived hemodynamic fields (velocity, pressure, wall shear stress) at multiple flow rates. Normalized pressure is computed as $P^* = \frac{P}{0.5\rho V_\text{max}^2}$, and simulation is performed using validated mesh and PISO solvers, supporting integrated morpho-hemodynamic learning (Li et al., 17 Jan 2025).

4. Feature Transfer, Unsupervised and Weakly Supervised Learning

Recognition of data scarcity and domain difference has motivated multiple approaches to feature learning and network design:

• Cross-Domain Feature Transfer: TRELLIS generates latent geometric descriptors from non-medical 3D objects and transfers them as 1024-dimensional feature embeddings for vascular surface classification, segmentation, and GNN-based blood-flow simulation. Integration of these features yields up to 95% accuracy improvement in challenging aneurysm discrimination tasks and 15% RMSE reduction in hemodynamic prediction (Hervé et al., 3 Sep 2025).
• Unsupervised Dual-Branch Learning: Unlabeled vessel segments are augmented by jittering; dual-branch (PointNet/PointNet++) encoders maximize mutual information between paired representations via a temperature-scaled contrastive loss:

$$l(i, j) = -\log \frac{\exp(s_{i, j}/\tau)}{\sum_k \exp(s_{i, k}/\tau)}$$

where $s_{i, j}$ is the cosine similarity, enabling high-performance transfer to supervised downstream classification/segmentation even with minimal labeled data (Shao et al., 2022).

• Weakly Supervised and Anatomically Guided Pretraining: MAE-style pretraining with 3D Vision Transformers extracts robust, vessel-centric features by masking and reconstructing both image intensities and artery distance maps from sub-volumes overlapping arterial domains. This approach, utilizing factorized slice-wise and axial self-attention, advances sensitivity by 4–8% at fixed false positive rates over prior methods, most markedly on out-of-distribution CT datasets (Ceballos-Arroyo et al., 28 Feb 2025). Similarly, weakly supervised multi-task UNet with vesselness attention blocks via Frangi filtering achieves Dice scores up to 0.614 and detection sensitivities above 92% (Rainville et al., 1 Aug 2025).

5. Localization, Landmark Detection, and Interpretation

Accurate anatomical localization remains crucial both for risk assessment and as a precursor to detailed segmentation:

• Bifurcation and Landmark Detection: The highest concentration of aneurysms occurs at major Circle of Willis bifurcations. Automated cascaded pipelines combine 3D U-Net vascular segmentation, skeletonization, and patchwise 3D CNN classification to identify 13–15 bifurcations of interest, yielding recognition rates >90% for most locations (Nader et al., 2023). Mean radial error for landmark localization is reduced to ~2 mm via two-stage detection and heatmap regression methods, with cascading architectures mitigating ambiguities in densely clustered bifurcations (Nader et al., 3 Jul 2025).
• Hybrid Heuristic–Learning Post-Processing: Anatomy-based post-processing, using automated artery/vein/brain/CVS masks, reduces false positive rates in CTA-based DL detection models by 70% without sensitivity loss. Overlap criteria (e.g., removing a prediction when the vein mask overlap exceeds arterial overlap) and targeted brain region constraints systematically remove non-aneurysm detections (Kim et al., 1 Jul 2025).
• Interpretability and Clinical Utility: These post-processing and modular approaches enhance model interpretability—allowing radiologists to corroborate AI-driven detections with familiar anatomical strata—and maintain confidence in automated support during both diagnosis and preoperative planning.

6. Clinical and Research Applications, Performance Metrics, and Limitations

Aneurysm detection frameworks are increasingly evaluated with clinically relevant metrics and deployed in multi-site studies:

• Performance Benchmarks: Sensitivity (typically >80–85%), false positive per case (down to ~0.2–0.4), Dice scores for segmentation ($\sim$0.73 for large unruptured IAs (Indrakanti et al., 30 Aug 2024)), and recognition rates at key bifurcations (>90% for large/central CoW segments (Nader et al., 2023)) are standard. Nodal-level Dice coefficients in surface segmentation (e.g., $0.88$ in GraphNet (Ma et al., 2019)) reflect spatial precision.
• Generalizability: Models are explicitly benchmarked across imaging modalities (TOF-MRA, CTA, DSA), with modality-agnostic solutions (such as geometric patch-based PointNet (Bizjak et al., 2020)) demonstrating consistent cross-domain sensitivity and low false positives. External validation (e.g., Lausanne/ADAM datasets (Rainville et al., 1 Aug 2025)) confirms robustness across institutions and scanners.
• Data Sharing and Open Source: Public release of model weights (e.g., nnU-Net at 10.5281/zenodo.13386859, AneuG meshes, and TRELLIS feature extractors) facilitates replication and further research advancement.

Limitations remain in reliably detecting extremely small or radiolucent aneurysms, and in handling anatomical variability, adjacent artifacts, and class imbalance. Synthetic augmentation, anatomically informed pretraining, and feature transfer strategies address these gaps and remain active areas of research.

7. Future Directions and Broader Impact

Emerging research directions center on:

• Automated full-pipeline reconstruction (e.g., from raw TOF-MRA to surface/volume): Integrating geometry extraction, detection, segmentation, and post-processing within end-to-end networks.
• Morpho-hemodynamic integration: Exploiting paired synthetic and real CFD data to link complex shape parameters with hemodynamic risk factors and simulating the effect of endovascular interventions (Li et al., 17 Jan 2025).
• Personalized, uncertainty-aware modeling: Leveraging MAE/Transformer architectures for fine-grained, interpretable prediction with uncertainty quantification and robust feature learning under distribution shifts (Ceballos-Arroyo et al., 28 Feb 2025).
• Fine-tuning and transfer learning: Cross-domain generative models trained on large non-medical corpora (e.g., TRELLIS (Hervé et al., 3 Sep 2025)) are being adapted for medical structures, dramatically improving data efficiency and downstream diagnosis.

A plausible implication is increasing adoption of such approaches in clinical workflow, providing fast, reproducible, and interpretable aneurysm detection, segmentation, and risk quantification support to neuroradiologists, and enabling informed preoperative assessment and follow-up. As large-scale annotated and synthetic datasets proliferate alongside open-source model repositories, continual benchmarking and refinement are expected to accelerate progress and broaden application to related vascular pathologies.