Connectivity-based Cerebrovascular Segmentation in Time-of-Flight Magnetic Resonance Angiography

Zan Chen, Xiao Yu, Yuanjing Feng

Abstract: Accurate segmentation of cerebrovascular structures from TOF-MRA is vital for treating cerebrovascular diseases. However, existing methods rely on voxel categorization, leading to discontinuities in fine vessel locations. We propose a connectivity-based cerebrovascular segmentation method that considers inter-voxel relationships to overcome this limitation. By modeling connectivity, we transform voxel classification into predicting inter-voxel connectivity. Given cerebrovascular structures' sparse and widely distributed nature, we employ sparse 3D Bi-level routing attention to reduce computational overhead while effectively capturing cerebrovascular features. To enhance directional information extraction, we utilize the 3D-direction excitation block. Additionally, the 3D-direction interactive block continuously augments direction information in the feature map and sends it to the skip connection. We compare our method with current state-of-the-art cerebrovascular segmentation techniques and classical medical image segmentation methods using clinical and open cerebrovascular datasets. Our method demonstrates superior performance, outperforming existing approaches. Ablation experiments further validate the effectiveness of our proposed method.

Primary Subject Area: [Experience] Multimedia Applications

Relevance To Conference:

Considering the propensity for voxel-based classification methods to cause disconnects in the thin cerebrovascular, we transformed the task of cerebrovascular segmentation from voxel classification to predicting inter-voxel connectivity.
Considering the beneficial hidden directional information within feature maps for connectivity prediction, we devised a 3D direction excitation block and a 3D direction interactive block to extract and facilitate the flow of hierarchical directional information between the encoder and decoder.
Our experimental results demonstrate superiority on clinical and public datasets compared with four state-of-the-art cerebrovascular segmentation models and two classic medical segmentation methods.

Submission Number: 2650