Efficient Repairing of Disconnected Pulmonary Tree Structures via Point-based Implicit Fields

Ziqiao Weng, Jiancheng Yang, Dongnan Liu, Weidong Cai

Keywords: Pulmonary Tree Segmentation, Pulmonary Tree Reconstruction, Lung Diseases, Point Cloud, Implicit Neural Network.

Abstract: Segmentation of pulmonary tree structures is critical for diagnosing and planning treatment for lung diseases. However, existing deep learning models often yield inaccurate segmentations, resulting in disconnected vessel predictions. To overcome this challenge, we propose an efficient framework for reconstructing pulmonary trees. Initially, we represent disconnected pulmonary tree structures as sparse surface point clouds. Next, we utilize a point cloud network to extract features and predict the disconnected segments. Finally, we employ an implicit neural network to infer the occupancy of arbitrary points, thereby facilitating efficient reconstruction. We validate the effectiveness of our approach on real data from 799 subjects; the code and data will be publicly available.

Submission Number: 74