Clinical Scorecard: Anatomy-Informed Masked Autoencoder with Domain-Adaptable Prompting for Multimodal Detection and Segmentation of Cerebral Aneurysms

At a Glance

Key Highlights

• AMAP framework improves detection and segmentation of cerebral aneurysms, especially small lesions under 5 mm.
• Incorporates anatomy-guided self-supervised learning and lesion-aware prompts to enhance localization and cross-domain robustness.
• Achieves 3–8% higher Dice scores and reduces false positives by ~20% across multiple public datasets and unseen domains.

Guideline-Based Recommendations

Diagnosis

• Utilize CTA and TOF-MRA imaging modalities for aneurysm detection and assessment.
• Apply anatomically-guided AI models like AMAP to improve sensitivity for small aneurysms and reduce false positives.

Management

• Incorporate AI-assisted detection and segmentation outputs to support risk stratification and treatment planning.
• Use prompt-guided visualizations to align AI findings with radiological practice for clinical decision-making.

Monitoring & Follow-up

• Validate AI model performance across multiple imaging centers and modalities to ensure domain generalization.
• Monitor false positive rates and calibration metrics to maintain diagnostic reliability.

Risks

• Be aware of potential reduced model performance on external cohorts without domain adaptation.
• Consider limitations in interpretability of foundation models; use anatomically-informed prompts to enhance clinical trust.

Patient & Prescribing Data

Patients undergoing neurovascular imaging for suspected cerebral aneurysms

AMAP enhances detection accuracy and segmentation quality, particularly for small aneurysms, facilitating improved clinical management and reducing diagnostic errors.

Clinical Best Practices

• Employ anatomically-guided pretraining to focus AI models on vascular structures for better aneurysm sensitivity.
• Use lesion-aware prompting to guide model attention to aneurysm-prone regions, improving localization with limited annotations.
• Incorporate boundary-aware contrastive learning to reduce false positives at vascular bifurcations and enhance domain robustness.
• Validate AI tools on multi-center datasets and unseen domains to ensure generalizability.
• Leverage prompt-based visualizations to improve interpretability and clinical acceptance of AI predictions.

References

• Intracranial cerebral aneurysms and subarachnoid hemorrhage risk
• Deep learning for aneurysm detection in CTA and TOF-MRA
• Vision Transformers and Masked Autoencoders in medical imaging
• Domain generalization challenges in medical AI
• Prompt learning for medical image analysis