부산대학교 도서관

Intracranial Aneurysm (IA) lesion segmentation is significant for IA treatment, which is one of the high death rate and deformity cerebrovascular diseases. Segmenting the IA lesions accurately is still challenging in digital subtraction angiography (DSA) images due to blurred boundaries, imaging noise, and intracranial vascular morphologies. In this paper, we are the first time to propose a novel instance segmentation network architecture, Graph Mask2Former, to segment IA lesions automatically based on DSA images. Specifically, we apply a graph convolution module to reassign label information, aiming to adjust the confidence weight of error instances adaptively. Furthermore, we design a local refinement module to refine the coarse mask output. The extensive experiments on the clinical IA- DSA and LiTS datasets show that our method outperforms recent state-of-the-art methods. This paper also provides the visual analysis to explain the inherent behavior of our method.