CREANet: a curvature retrospective and edge-aware network for precise intracranial aneurysm segmentation in point clouds

Intracranial aneurysm refers to the abnormal expansion of the cerebral artery wall. When it ruptures, it can lead to subarachnoid hemorrhage, threatening the patient’s health. In this work, we focus on segmenting intracranial aneurysms from point clouds reconstructed from 3D Time-of-Flight Magnetic Resonance Angiography volumes. Due to the complex geometric structure and blurred boundaries of aneurysms, the segmentation task has become highly challenging. To improve the segmentation performance, we proposed a curvature retrospective and edge-aware network (CREANet), which integrates curvature traversal, retrospective optimization, and edge-aware enhancement for precise aneurysm segmentation. This network employs a point recombination strategy based on curvature features for spectral analysis to enhance aneurysm geometry extraction. It also utilizes a multilevel recursive feature optimization module to preserve structural details, thereby achieving the effective flow of information. Furthermore, the proposed edge-aware enhancement module can improve edge recognition performance by predicting edge probability. In the IntrA dataset, CREANet achieves an IoU of 95.06% for vessel segmentation and 84.42% for aneurysm segmentation, while CREANet attains an A_ACC of 94.56% in the classification task, outperforming some state-of-the-art techniques. In addition, CREANet further validates its robustness and generality on the ModelNet40 and ShapeNetPart datasets with 93.4% overall accuracy and 84.7% mIoU, respectively. In summary, CREANet offers an efficient solution, improving the performance of IAs segmentation.