Abstract
Deformable shapes play a pivotal role in numerous image analysis tasks, such as image registration for real-time image-guided navigation systems in tumor removal surgery, image classification to detect neuro-degenerative diseases, and template-based image segmentation for object tracking. In recent years, the advancements in deep learning-based image analysis have resulted in remarkable performance by providing a universal mechanism for extracting image features in the context of texture, intensity, or simple geometry features that may be hidden behind raw data. Nevertheless, existing deep learning methods fail to recognize geometric deformable shape features that can capture complex and detailed geometric information in images, resulting in a considerable limitation of the image analysis models when the analysis and quantification of geometric shapes are crucial.
Learning and modeling deformable shapes is particularly challenging due to their high dimensional and non-linear nature of data, which inevitably cause expensive network training and inference with high computational complexity. In addition, current related algorithms suffer from time and labor-consuming parameter tuning and appearance change (i.e., caused by missing data values, corrupted signals, or occurrence of objects). To address these issues, the ultimate goal of my dissertation is to develop a deep learning framework that learns and analyzes efficient and robust representations of shapes central to image analysis tasks. This research naturally merges low-dimensional deformable geometric shape features in various deep learning based image analysis tasks, including but not limited to image registration, image classification, image segmentation, uncertainty quantification, atlas building, and parameter estimation. The developed framework has great impact on a variety of real-world clinical applications. For example, it allows neurosurgeons to identify brain shifts (deformations) caused by multiple factors during surgery (i.e., gravity, fluid drainage, or changes in intracranial pressure and swelling of brain tissue), and modify surgical plans in real-time. Such a framework helps clinicians better understand, interpret, and analyze the image registration results and further enables precise diagnosis according to the patient's clinical condition. This research also facilitates more robust clinical diagnostic routines in neuro-degenerative disease prediction, such as Alzheimer's disease detection, and post-treatment for patient care.
In particular, I first develop efficient image analysis models based on low dimensional shape representations. The outcome of my research speeds up Bayesian uncertainty quantification for image registration, enable fast atlas building with automatic parameter selection, and allow rapid training data generation for learning-based registration regularization parameter estimation. I then develop deep neural networks to learn low-dimensional shape representations through image registration with much lower computational complexity in training. To further enhance the efficiency of the image analysis models, I further utilize learned deformable shape representations for population-based image studies by developing joint image classification models with group mean estimation, also known as atlas building, which improves model accuracy, robustness, and efficiency. Additionally, I introduce a deep metamorphic neural network that effectively controls geometric shape deformation and appearance changes (i.e., caused by tumor resection), leading to precise image registration with lower model error. My dissertation has significant potential to impact clinical applications, such as automated diagnosis for neuro-degenerative diseases and image-guided navigation systems for neurosurgery. Overall, my research provides a theoretical foundation in machine learning and computer vision, facilitating better image analysis in unmet clinical needs and neuroscience studies.
