Graphical methods for atlas-based segmentation and local disease grading in medical images

Abstract

Advances in medical imaging offer valuable tools towards a better understanding of the human anatomy and the diversity of physiological and pathological processes present in the body. Image segmentation is a crucial building block in medical image analysis, allowing quantitative analysis of the image content. It is therefore used as a fundamental step in a myriad of clinical applications. As the wealth of available data increases, automating this process is imperative as manual segmentation is very time consuming and may require high levels of expertise. Automated segmentation approaches face challenges in large databases due to large variability in shape and appearance of the structures of interest, the presence of pathologies, or different imaging protocols used to acquire the images. In particular, it becomes increasingly desirable to develop robust and accurate segmentation techniques that rely on minimal manual input or weak supervision. The contributions presented in this thesis are three-fold: We present two graphical methods for atlas-based anatomical image segmentation with a focus on alleviating the impact of limited availability of manually annotated training data. In addition, we present a probabilistic anatomical model based on segmentation which can help analyse regional morphological differences between groups, e.g. between healthy subjects and dementia patients, and provide local disease scores in the images which express the estimated degree of pathology. The proposed segmentation methods were evaluated on segmentation tasks on fetal and adult brain MR images and adult cardiac MR images. Experiments demonstrated the flexibility of the proposed approaches and showed that accurate segmentation results could be obtained while reducing the amount of manually annotated input. The proposed anatomical disease model was evaluated on a clinical brain MR dataset for differentiation between Alzheimer’s disease patients and healthy elderly subjects. Experiments showed that modelling disease-specific anatomy in higher detail revealed a greater sensitivity to local anatomical differences between the studied groups.