Deep Learning Segmentation Algorithms for X-ray CT data

The segmentation task for 3D objects from X-ray CT volumetric data is of great significance for both industrial and medical applications. Deep learning techniques are narrowing the gap between human and machine capabilities in image segmentation. In this thesis we develop and discuss machine and deep learning techniques for semantic and instance segmentation. The techniques are evaluated on a dataset of CT scans of short glass fiber reinforced polymers prepared in cooperation with the University of Padova and on publicly available medical CT scans of lungs and liver. In addition to that, the last chapter is evaluated on a public and popular large-scale object detection, segmentation, and captioning dataset for a better comparison with the state-of-the-art. The chapters are structured in the following way: In chapter 2 we explain the short glass fiber reinforced polymer data acquisition together with the reference setup for quantitative comparison of segmentation techniques. The data creation process involves parts manufacturing, CT scanning, CT simulation, computational model design, volume reconstruction and ground-truth preparation. The reference setup consist of metrics for instance and semantic segmentation tasks as well as of a baseline, Frangi vesselness method. In chapter 3 we present a first deep learning model for semantic segmentation of fibers from CT scans. The model outperforms all the other methods including feature-engineered and machine learning models. In chapter 4 we present a first deep learning model for instance segmentation of fibers from CT scans. The model outperforms the state-of-the-art by a significant margin and is arguably the first method which allows calculation of important fiber statistics based on single-fiber segmentation. The model consist of a fully convolutional branch for semantic segmentation, and an enhanced branch for instance segmentation via proposed embedding learning loss function. In chapter 5 we present our work on use of machine learning techniques for medical CT analysis. We use a dictionary learning model and extend it to a 3D for bronchial vessels segmentation from thorax CT scans. Then, we discuss and develop a fully convolutional deep learning model for the task of liver and liver lesion segmentation from liver CT scans. Lastly, we present the Mask Mining training approach for boosting the semantic segmentation machine learning models. In chapter 6 we present the idea of the Plugin Networks as a solution for inference under partial evidence. The proposed framework can generalize to a number of machine learning tasks and is evaluated on the task of hierarchical scene categorization, multi-label image annotation and scene semantic segmentation achieving state-of-the-art on each