Date of Award

12-2026

Document Type

Thesis

Degree Name

Master of Science (MS)

Department

Electrical Engineering and Computer Science

First Advisor

Neji Mensi

Second Advisor

Michael King

Third Advisor

Jignya Patel

Fourth Advisor

Brian A. Lail

Abstract

Many publicly available cardiac datasets were developed to address specific clinical or research objectives and therefore provide annotations for only a subset of cardiac structures. Although fully annotated datasets exist, a large amount of partially labeled cardiac data remains underutilized because existing reconstruction methods typically require complete whole-heart anatomy. This thesis proposes a template-guided graph neural network for reconstructing missing cardiac structures from partially segmented anatomy. The available structures are represented as point clouds, encoded using a shared PointNet-based encoder, and combined with a fixed anatomical template through a GraphSAGE-based graph neural network to predict the missing structures. The proposed framework is evaluated on the MMWHS, ImageCHD, HVSMR-2.0, and ACDC datasets using geometric reconstruction metrics. Experiments investigate reconstruction accuracy, the contribution of the anatomical template, patient-specific conditioning, cross-dataset generalization, and reconstruction from arbitrary subsets of known cardiac structures. Results demonstrate that the anatomical template substantially improves reconstruction accuracy, the framework generalizes to previously unseen datasets without retraining, and reconstruction performance depends more strongly on the identity of the available anatomical structures than on their number. These findings demonstrate that template-guided graph neural networks provide a practical approach for reconstructing complete whole-heart anatomy from incomplete cardiac annotations.

Download

Available for download on Tuesday, June 01, 2027

Share

COinS