Two-Steps Neural Networks for an Automated Cerebrovascular Landmark Detection
This study addresses the challenge of accurately and robustly localizing 13 critical arterial bifurcations in the Circle of Willis (CoW) for intracranial aneurysm (ICA) diagnosis. We propose a two-stage deep learning framework: first, a detection network coarsely identifies bifurcation candidates; second, a deeply supervised, modified U-Net performs sub-pixel-level segmentation and precise localization. This design effectively mitigates bifurcation omission due to spatial proximity and enhances generalizability across anatomical CoW variants. Evaluated on both a private and public MRA dataset, our method significantly outperforms existing single-stage models. It achieves high accuracy across diverse vascular morphologies and bifurcation counts—yielding a mean localization error < 0.8 mm and recall > 97.5%. These results demonstrate strong clinical applicability and methodological advancement in automated CoW bifurcation localization.