Hybrid-scale fusion network improves intracranial aneurysm segmentation on CTA

A Swin UNETR-based deep learning network with hybrid-scale fusion and attention refinement achieved mean patient-level Dice of 83.4% for intracranial aneurysm segmentation on CTA, outperforming baseline and showing preliminary cross-center applicability.

• HSF-Net uses a "fuse first, refine later" strategy, integrating hybrid-scale fusion and convolutional block attention modules into a Swin UNETR decoder.
• On the internal test set, it achieved an 83.4% mean Dice score, and an independent external cohort provided preliminary evidence of cross-center generalization.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.