Self-Supervised Representation Learning in Point Clouds for Hierarchical Graph-Based Anatomical Structure Identification

Rocholl, Niels (2024) Self-Supervised Representation Learning in Point Clouds for Hierarchical Graph-Based Anatomical Structure Identification. Master's Thesis / Essay, Artificial Intelligence.

Preview	Text mAI2024RochollN.pdf Download (59MB) | Preview
	Text toestemming.pdf Restricted to Registered users only Download (129kB)

Abstract

This thesis presents Graph Representation of Advanced Part Encodings (GRAPE), a novel framework for anatomical structure identification in three-dimensional point clouds of human bodies. By combining self-supervised learning and graph neural networks, GRAPE constructs graph-based representations of 3D objects, where each node is enriched with a latent feature derived from their corresponding object parts. These features, generated by a Masked Autoencoder (Point-MAE), capture geometrical and spatial information within the point cloud. GRAPE-GNN, the model developed using this framework, leverages both the node embeddings (latent Point-MAE features) and the topological information within the hierarchical graph created by GRAPE for accurate identification of anatomical structures. Evaluated on our Anatomy-GRAPE dataset, which was specially annotated for this thesis and validated by medical experts, GRAPEGNN demonstrates high precision and recall across various anatomical structures. Our experiments show that GRAPE-GNN retains good performance in scenarios with limited annotated training data. We identify areas for improvement, such as enhancing performance under point cloud occlusion. This method not only demonstrates the effectiveness of graph-based anatomical structure identification but also the potential of integrating self-supervised learning based point cloud features with graph-based representations in general.

Item Type:	Thesis (Master's Thesis / Essay)
Supervisor name:	Sabatelli, M.
Degree programme:	Artificial Intelligence
Thesis type:	Master's Thesis / Essay
Language:	English
Date Deposited:	30 Aug 2024 11:28
Last Modified:	30 Aug 2024 11:28
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/34128

Actions (login required)

View Item