Encoding anatomical tree priors for tubular structure extraction for medical images analysis

Vasculatures and airways in the human body contain anatomical trees, which are tree structures with anatomical properties and geometrical attributes. Since anatomical trees are highly involved in clinical procedures such as disease diagnosis and treatment planning, accurate and automatic annotation and analysis of these structures is extremely important. In this thesis, after an extensive study of existing literature on 3D tubular tree analysis, we introduce novel techniques encoding anatomical tree priors into vasculature and airway extraction. We present novel features, e.g., features based on multi-modal von Mises-Fisher distribution, for 3D vasculature bifurcation classification using Random Forest classifier. Then we introduce the first work fitting a parametric 3D geometric model to 3D medical image data of pulmonary vasculature for bifurcation localization. To solve the corresponding optimization problem, we present the modified genetic algorithm with tribes niching technique. For encoding the geometrical variability of anatomical trees and their natural sequential root-to-leaf representation, we propose two deep learning models, the TreeNet and LSTM-Tree for predicting branch direction and bifurcation classification during centerline tree tracking. To overcome the myopic visual search involved in most tree tracking processes, we introduce two novel ways to leverage global prior information, by using tree-level statistics within a Bayesian framework and reframing the tree shape into a pictorial structure. Then we encode anatomical tree priors in the clinical task of age-related macular degeneration classification and retinopathy grading by masking a sequential attention within deep network layers.