SIAM: Spatial and Intensity Awareness Module for cerebrovascular segmentation

Cerebrovascular segmentation plays a crucial role in guiding the diagnosis and treatment of cerebrovascular diseases. With the rapid advancements in deep learning models, significant progress has been made in 3D cerebrovascular segmentation. However, they often rely on massive images and annotations, which is still challenging in cerebrovascular segmentation.

Considering the unique pixel and spatial features inherent to vascular structures, such as vessel shape, location, and high pixel intensity characteristics, we propose a novel Spatial and Intensity Awareness Module (SIAM) for limited cerebrovascular segmentation. This module introduces spatial and pixel intensity perturbations to construct new matching data for model learning. Using collaborative training and shared features, SIAM gains the awareness of spatial and pixel intensity, thereby endowing the model with cerebrovascular semantics. Owing to the awareness learning belonging to an independent training module, SIAM satisfies the attribute of plug-and-play.

To validate SIAM, we carried out experiments on three cerebrovascular datasets with different modalities. The results demonstrate that SIAM enables the models to perform remarkably in normal and limited cerebrovascular segmentation. It can be seamlessly integrated into existing segmentation models without disrupting structural integrity.

SIAM effectively learns and adapts to the unique spatial and pixel intensity features of vascular structures through collaborative training and shared features. Our experiments on three different cerebrovascular datasets confirm its robustness and efficacy even with limited data. Furthermore, its plug-and-play nature allows for seamless integration into existing models, preserving their structural integrity. Our code is available at https://github.com/QingYunA/SIAM-Spatial-and-Intensity-Awareness-Module-for-3D-Cerebrovascular-Segmentation.