Deformable Convolution and Semi-supervised Learning in Point Clouds for Aneurysm Classification and Segmentation

Abstract

Intracranial aneurysm is a life-threatening and high-risk abnormality. In practice, both aneurysm classification and segmentation are very important for diagnosis and treatment planning. There have been various studies of automatic diagnosis of aneurysms based on medical images with 2D image processing methods. However, the diagnosis of an aneurysm based on 3D models is potentially much more accurate than on 2D images. The edge of an aneurysm is much clearer with 3D visualization, and the complicated and time-consuming annotation on 2D images can be avoided. Here we propose a deep neural network incorporating deformable point convolution and self-attention to improve the classification and segmentation performance for 3D point clouds. In addition, we introduce semi-supervised learning to address the problem of class imbalance. Our experimental results on a public 3D point clouds dataset show that our model outperforms several state-of-the-art deep learning models.