Robust topology-preserving segmentation in medical images
File(s)
Author(s)
Li, Liu
Type
Thesis
Abstract
Segmentation is an important step in medical image analysis that offers quantifiable insights to clinicians. Despite high pixel-wise accuracy for specific segmentation tasks, real-world challenges including imaging artefacts, insufficient labels, and poor generalisation ability still cause segmentation errors. This thesis focuses on topological errors where the anatomical structure of the segmentation differs from the medical prior, e.g., discontinuous boundaries and broken surfaces.
To address these issues, we propose four approaches that utilise shape and topology priors. The motivation is that despite the shape variance of organs across different samples, they share a common anatomical structure. This thesis progresses from addressing specific challenges in fetal Magnetic Resonance (MR) images to developing general topological refinement solutions that apply to diverse applications.
We start with fetal brain segmentation using atlas-based shape priors (Chapter 3) and Persistent Homology (PH)-based topology priors (Chapter 4). In Chapter 5, we extend this to general topology-aware segmentation tasks with Euler Characteristic (EC). Chapter 6 further enhances the robustness using synthetic data generated from polynomial bases.
Initially, we develop CAS-Net, a novel network that integrates shape priors via conditional atlases. This method is demonstrated on the dHCP dataset for fetal brain segmentation with a high Dice Similarity Coefficient (DSC) score.
Building on shape priors, we introduce topological priors for Cortical Grey Matter (CGM) segmentation, proposing a weakly-supervised method that learns from noisy labels with CGM thickness and PH regularisation. This demonstrates the effectiveness of priors under morphological variability and label limitations.
Furthermore, we introduce a general and efficient EC-based topology-aware segmentation approach that detects and refines topology violations, improving pixel-wise accuracy and topological integrity across datasets.
Finally, we propose a model-agnostic topology refinement network trained with diverse synthetic errors using orthogonal polynomial bases. This method allows an efficient refinement without model-specific retraining.
To address these issues, we propose four approaches that utilise shape and topology priors. The motivation is that despite the shape variance of organs across different samples, they share a common anatomical structure. This thesis progresses from addressing specific challenges in fetal Magnetic Resonance (MR) images to developing general topological refinement solutions that apply to diverse applications.
We start with fetal brain segmentation using atlas-based shape priors (Chapter 3) and Persistent Homology (PH)-based topology priors (Chapter 4). In Chapter 5, we extend this to general topology-aware segmentation tasks with Euler Characteristic (EC). Chapter 6 further enhances the robustness using synthetic data generated from polynomial bases.
Initially, we develop CAS-Net, a novel network that integrates shape priors via conditional atlases. This method is demonstrated on the dHCP dataset for fetal brain segmentation with a high Dice Similarity Coefficient (DSC) score.
Building on shape priors, we introduce topological priors for Cortical Grey Matter (CGM) segmentation, proposing a weakly-supervised method that learns from noisy labels with CGM thickness and PH regularisation. This demonstrates the effectiveness of priors under morphological variability and label limitations.
Furthermore, we introduce a general and efficient EC-based topology-aware segmentation approach that detects and refines topology violations, improving pixel-wise accuracy and topological integrity across datasets.
Finally, we propose a model-agnostic topology refinement network trained with diverse synthetic errors using orthogonal polynomial bases. This method allows an efficient refinement without model-specific retraining.
Version
Open Access
Date Issued
2024-10-08
Date Awarded
01/01/2025
License URL
Advisor
Rueckert, Daniel
Kainz, Bernhard
Sponsor
Lee Family Scholarship
Publisher Department
Computing
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)