Contrastive Learning with Self-reconstruction for 3D Intracranial Aneurysm Detection

Medical point clouds have rich 3D geometric information for intelligent disease detection. However, it is usually time-consuming and costly to annotate such medical data. In this paper, we propose a novel self-supervised method for intracranial aneurysm detection based on 3D medical point cloud data. In particular, we devise a dual-branch network which consists of contrastive learning and reconstruction task, namely CORE. Our method enables the effective extraction of both integral and point-wise features of point clouds. We specifically design a two-part reconstruction decoder. The first part is a fully-connected decoder to predict a coarse point set that represents the global geometry. The second part is a folding-based decoder to approximate the surface of the local geometry. Then we employ a dual-optimiser strategy so that the parameters of each branch are optimised separately to avoid the interaction between contrastive learning and reconstruction that may cause the local optimum issue. We conduct two evaluation schemes on the public IntrA dataset: self-supervised linear classification and fine-tuning. Experiments demonstrate CORE achieves state-of-the-art performance in terms of both evaluation schemes on 3D intracranial aneurysm classification and segmentation. In addition, the generalization ability of CORE is also successfully verified on ModelNet40/10 and ShapeNetPart.