Ehancing Robustness in Aneurysm Segmentation Across Variable Sizes via ContextAware Semantic Augmentation

Introduction: Intracranial aneurysms (IAs) are vascular abnormalities that can lead to life-threatening hemorrhage upon rupture. Accurate segmentation on 3D Time-of-Flight Magnetic Resonance Angiography (TOFMRA) is essential for diagnosis and treatment planning, but remains difficult owing to variability in IA size and appearance. Particularly, small or low-contrast aneurysms are frequently missed by deep learning-based models.

Data augmentation is commonly used to improve data diversity, but conventional methods focus on low-level perturbations and often fail to reflect clinically meaningful variations. To address this, we propose a context-aware semantic augmentation (CSA) strategy that modulates intensity patterns while preserving anatomical structure. CSA introduces realistic, size-sensitive variability into training data, aiming to enhance robustness in segmenting IAs of diverse sizes, especially small or subtle lesions.

Methods: We collected 669 3D TOF-MRA scans from two institutions and split them into training and validation sets in a 3:1 ratio. For model input, 3D vascular patches were extracted based on segmented vessel maps, where threshold values were determined through statistical comparison between Gumbel and unilateral normal distributions. Vessel skeletonization was then applied to identify centerlines and flow directions for patch sampling.

During training, our CSA manipulates intensity distributions in a structure-preserving manner using a Bezier-based contrast transformation. It comprises two stages: (1) global location scaling, which adjusts contrast across the full volume, and (2) local location scaling, which targets specific regions indicated by the ground truth mask. These are blended using a spatially varying tile map ranging from 0 to 1, generating augmented inputs that simulate plausible variations in IA appearance.

The segmentation model, SwinUNETR, was trained using these CSA-enhanced patches and corresponding ground truth masks, optimized with a combination of Dice and cross-entropy losses.

Results: The model was thoroughly evaluated on an external test set comprising 209 3D TOF-MRA scans. Quantitative metrics, including Dice Similarity Coefficient (DSC), Precision, and Recall, were computed to assess segmentation accuracy. Compared to conventional geometry and intensity augmentations, CSA preserved vessel morphology while introducing subtle, diverse intensity variations. As a result, CSA significantly outperformed all baselines, achieving scores of DSC: 50.30%, Precision: 53.81%, and Recall: 52.12% across IA size categories.

Discussion & Conclusion: We proposed a CSA to enhance IA segmentation robustness across variable sizes. CSA introduces anatomically consistent intensity variations, leading to improved detection capabilities of small or subtle IAs. Beyond accuracy, CSA contributes to more reliable volume estimation and supports clinical decision-making, such that it offers a practical solution for scalable and real-world deployment in neurovascular imaging.