REGION-BASED LANDMARK DETECTION AND STRUCTURE-PRESERVING SEGMENTATION IN MEDICAL IMAGES

Anatomical landmark detection and medical image segmentation are two basic and long-standing problems in medical image analysis. They play essential roles in pre-operative assessment, morphological analysis, disease diagnosis, anatomical understanding, treatment planning, and image-guided intervention. Due to the lack of reliable methods for anatomical landmark detection and image segmentation, the current clinical practice is still largely based on manual and semi-automatic annotation, requiring a significant amount of human intervention, which is time-consuming, error-prone, experience dependent, and susceptible to intra- and inter-observer variability. Therefore, computational tools to automate anatomical landmark detection and medical image segmentation are highly desired. Existing methods for these tasks can be in general divided into two categories: traditional and deep neural network (DNN) methods. Traditional methods can be further divided into atlas-based, model-based (e.g., deformable models and level-set models), and traditional machine learning (ML) methods. However, atlas-based methods are often computationally intensive and very sensitive to large shape variations and imaging artifacts. Model-based methods require setting up an initial model and manually choosing appropriate parameters. Traditional ML-based methods rely heavily on the quality of hand-crafted features and thus require significant domain knowledge for feature engineering and feature selection. Due to the limited representation power of hand-crafted features and the independence of feature extraction and model training, traditional ML methods are less accurate and less robust in comparison to DNN-based methods. Recently, DNNs have demonstrated performance superior to traditional methods in anatomical landmark detection and medical image segmentation due to their ability in extracting hierarchical and representative features from training data. In this thesis, we focus on developing DNNs that aim at tackling challenging issues with existing methods for anatomical landmark detection and medical image segmentation. Our goal is to equip clinicians computational tools that can perform robustly under various conditions to significantly reduce the need for human intervention.