AMAP: Anatomy-Informed Autoencoder for Robust Cerebral Aneurysm Detection and Segmentation
Overview
AMAP integrates anatomy-guided masked autoencoder pretraining, domain-adaptive prompting, and boundary-aware contrastive learning to improve detection and segmentation of cerebral aneurysms. It achieves 3−5% higher Dice scores and reduces false positives by approximately 20% across multiple datasets and imaging modalities.
Background
Intracranial cerebral aneurysms are critical vascular abnormalities with a high risk of rupture, often leading to subarachnoid hemorrhage and death. Early and precise detection, especially of small aneurysms under 5 mm, is essential for risk assessment and treatment planning. Current AI methods struggle with small lesion detection, mis-segmentation at vascular bifurcations, and domain generalization across imaging centers and modalities. Advances in deep learning, including CNNs, transformers, and self-supervised learning, have improved aneurysm analysis but challenges remain in robustness and clinical interpretability.
Data Highlights
Metric
Improvement with AMAP
Dice Score
3−5% higher than baselines
False Positives per Case
~20% reduction
Datasets Evaluated
ADAM, IntrA, CQ500, plus unseen domains
Key Findings
Anatomy-guided MAE pretraining focuses self-supervised learning on cerebrovascular structures, enhancing sensitivity to subtle aneurysm morphology.
Domain-adaptive prompting combines global vascular priors with lesion-aware prompts, improving localization and cross-domain robustness.
Boundary-aware contrastive learning with GS-EMA optimization reduces false positives at vascular bifurcations and aligns vessel boundaries.
AMAP outperforms CNN, Transformer, and foundation-based models on multiple public datasets, achieving higher Dice scores and fewer false positives.
Qualitative results demonstrate accurate boundary preservation and consistent detection of small aneurysms often missed by other methods.
Clinical Implications
AMAP's improved detection and segmentation accuracy, particularly for small aneurysms, supports more reliable risk assessment and treatment planning. Its domain adaptability enhances performance across different imaging centers and modalities, facilitating broader clinical applicability. The prompt-guided visualizations offer interpretable outputs aligning with radiological practice, potentially increasing clinician trust in AI-assisted aneurysm screening.
Conclusion
AMAP represents a significant advancement in automated cerebral aneurysm analysis by integrating anatomical guidance, domain-adaptive prompting, and boundary-aware learning. This framework improves detection accuracy, reduces false positives, and enhances robustness, moving AI closer to trustworthy clinical deployment for aneurysm screening.
References
AMAP Study 2024 -- Anatomy-Informed Masked Autoencoder with Domain-Adaptable Prompting for Multimodal Detection and Segmentation of Cerebral Aneurysms
