Edge-Oriented Point-Cloud Transformer for 3D Intracranial Aneurysm Segmentation

Yifan Liu, Jie Liu, Yixuan Yuan

Abstract: Point-based 3D intracranial aneurysm segmentation is fundamental for automatic aneurysm diagnosis. Though impressive performances, existing point-based 3D segmentation frameworks still perform poorly around the edge between vessels and aneurysms, which is extremely harmful for the clipping surgery process. To address the issue, we propose an Edge-oriented Point-cloud Transformer Network (EPT-Net) to produce precise segmentation predictions. The framework consists of three paradigms, i.e., dual stream transformer (DST), outer-edge context dissimilation (OCD) and inner-edge hard-sample excavation (IHE). In DST, a dual stream transformer is proposed to jointly optimize the semantics stream and the edge stream, where the latter imposes more supervision around the edge and help the semantics stream produce sharper boundaries. In OCD, aiming to refine features outside the edge, an edge-separation graph is constructed where connections across the edge are prohibited, thereby dissimilating contexts of points belonging to different categories. Upon that, graph convolution is performed to refine the confusing features via information exchange with dissimilated contexts. In IHE, to further refine features inside the edge, triplets (i.e. anchor, positive and negative) are built up around the edge, and contrastive learning is employed. Differently from previous contrastive methods of point clouds, we only select points nearby the edge as hard-negatives, providing informative clues for discriminative feature learning. Extensive experiments on the 3D intracranial aneurysm dataset IntrA demonstrate the superiority of our EPT-Net compared with state-of-the-art methods. Code is available at https://github.com/CityU-AIM-Group/EPT .