Segmentation of cerebral vessels and aneurysms is vital for diagnosing cerebrovascular conditions, developing treatment plans, and supporting in silico trials.
The intricate nature of cerebrovascular anatomy and the small and irregular characteristics of aneurysms pose significant challenges for obtaining high-precision segmentation, which is vital for accurate clinical evaluations and planning of interventions.
The motivation stems from the necessity to address several critical challenges that compromise the efficiency of existing segmentation models. These challenges include class imbalance, where smaller structures, like aneurysms, are often under-represented in datasets, leading to suboptimal segmentation performance. Moreover, the insufficiency of labelled data constrains the potential of fully supervised models. In addition, the issue of domain shifts across various imaging modalities and patient populations can cause models trained on one dataset to perform inadequately on another. Finally, there is a need for models that can generalise effectively across diverse datasets from various clinical data sources, ensuring consistent performance regardless of the data source.
In response to these challenges, this thesis presents several novel contributions. It introduces a 3D patch-based multi-class model that effectively manages class imbalance and inter-class interference in vessel and aneurysm segmentation, employing advanced network elements and paired preprocessing and postprocessing techniques to enhance accuracy. A semi-supervised learning approach is also developed to leverage both labelled and unlabelled data, significantly improving segmentation consistency and continuity, particularly in scenarios with limited annotations. Furthermore, the thesis proposes a transwarp contrastive learning framework for unsupervised domain adaptation, allowing the model to handle domain shifts and perform robustly across different data modalities. Finally, a gradient-based domain generalisation method is introduced to ensure that segmentation models can generalise well across various imaging conditions, overcoming the challenges posed by data source variability.
