TAF-Net: Temporal-Adaptive Fusion Framework for Semisupervised Segmentation of Intracranial Arteries in DSA Sequences

Area of Science:

• Medical imaging analysis
• Deep learning for medical image segmentation
• Cerebrovascular imaging

Background:

• Accurate segmentation of intracranial arteries in digital subtraction angiography (DSA) is vital for diagnosing cerebrovascular conditions.
• Challenges include limited annotations and complex vascular structures, hindering precise analysis.
• Existing methods struggle with interframe inconsistency and vessel discontinuity in DSA sequences.

Purpose of the Study:

• To introduce TAF-Net, a semisupervised dual-path framework for enhanced intracranial artery segmentation in DSA.
• To leverage anatomical priors from MedSAM and fine-grained features from UNet for improved segmentation accuracy.
• To address challenges of interframe inconsistency and topological discontinuity in multiframe DSA sequences.

Main Methods:

• Proposed TAF-Net, a semisupervised dual-path framework integrating MedSAM and UNet.
• Introduced a temporal-adaptive fusion (TAF) strategy for dynamic prediction fusion based on confidence and temporal priors.
• Developed a spatiotemporal topology-aware loss to enforce structural continuity across frames.

Main Results:

• TAF-Net demonstrated superior performance on two public multiframe DSA datasets (DIAS and DSCA).
• Achieved state-of-the-art results in overlap accuracy (DSC, IoU) and topological integrity (Cost, 95HD).
• Showed significant improvements, particularly in low-label regimes and for maintaining vessel continuity.

Conclusions:

• TAF-Net effectively segments intracranial arteries in DSA, outperforming existing methods.
• The temporal-adaptive fusion and topology-aware loss are key to improving accuracy and structural integrity.
• This framework offers a promising solution for robust cerebrovascular diagnosis using DSA imaging.