Application of machine learning methods to the analysis of x-ray angiography images

Thesis

Invasive coronary angiography (ICA) is the gold standard imaging modality for diagnosing Coronary Artery Disease (CAD) during cardiac interventions. Accurate segmentation of coronary vessels in ICA could be beneficial for aiding diagnosis and developing effective treatment plans. However, automated vessel segmentation faces multiple challenges, including data scarcity, motion artifacts, uneven contrast distribution, and insufficient feature extraction for semantic analysis. To address these challenges, I first propose a semi-supervised segmentation framework based on a mean teacher model, employing Nested UNets as the backbone. The framework leverages unlabeled ICA images to extract informative features, while an elastic interaction-based loss function helps preserve structural integrity. This approach is trained and evaluated on a dataset collected from the Oxford John Radcliffe (JR) Hospital and demonstrated superior performance compared to state-of-the-art methods. Next, I present a novel Temporal Vessel Segmentation Network (TVS-Net), which fuses sequential ICA frames using a densely connected three-dimensional (3D) encoder and two-dimensional (2D) decoder structure, explicitly disentangling overlapping vessels and analyzing motion in ICA. The model is trained on an ICA dataset comprising 323 samples obtained from the Renji Hospital of Shanghai Jiao Tong University (SJTU), with additional out-of-distribution evaluation conducted on the dataset collected from the Oxford JR Hospital. The results demonstrate superior generalizability compared to six state-of-the-art methods. Finally, I explore both fusion and cascaded approaches by integrating convolutional neural networks (CNNs) with graph neural networks (GNNs) to incorporate geometric features. The proposed method enhances semantic vessel segmentation on a dataset I specifically constructed, using a novel graph generation algorithm and a node label penalty loss, achieving state-of-the-art performance across major coronary vessel branches in the selected view. All proposed methods demonstrate their effectiveness in addressing key challenges, consistently outperforming established benchmarks across multiple evaluation metrics. These advancements underscore the robustness and superiority of my methods for coronary vessel segmentation in ICA images.

Authors: He, H

• DOI: 10.5287/ora-eqpg2n7my
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: temporal information; vessel connectivity; nested encoder decoder; x-ray coronary angiography; coronary vessels segmentation
• Subjects: Coronary arteries; Biomedical engineering; Neural networks (Computer science); Image segmentation