N-PointNet: A multi-layer embedded deep learning model for 3D intracranial aneurysm classification and segmentation

Abstract Background In computer-aided intracranial aneurysm (IA) classification and segmentation, applications of 3D point cloud algorithms are increasingly widespread. However, the traditional point-based deep learning algorithm has the problem of poor segmentation effect. Methods An improved end-to-end depth network structure (N-PointNet) is proposed for IA classification and segmentation. First, the point cloud data of the IA are preprocessed. Then, the PointNet + + network structure is used as a backbone with learned hierarchical properties. After that, the preprocessed and resampled data produce multiple layers of information embedded in the original network input to further enhance its characteristics. Finally, a side output block is added, and the loss function of the corresponding layer is calculated. The multi-loss function facilitates fast convergence and improves model performance. Conclusion An experiment on the IntrA dataset proved the superiority of N-PointNet and obtained the best classification and segmentation results among the models tested. In addition, the proposed method has good generalization ability and has been verified on the common ModelNet40 dataset.