Barbano, Riccardo; (2024) Scalable Uncertainty Quantification and Learning for Deep Computed Tomography Reconstruction. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text Thesis_Scalable_Uncertainty_Quantification_and_Learning_for_Deep_Computed_Tomography_Reconstruction_riccardo_barbano.pdf - Accepted Version Download (21MB) | Preview

Abstract

Over the past decade, significant strides have been made in the field of medical image reconstruction due to the rapid progress of deep learning. This progress has been made possible through the integration of model-based imaging strategies with advances in large-scale neural architectures. Despite these achievements, challenges remain in utilising data-driven methods for image reconstruction in medical diagnostics. Challenges encompass poor generalizability properties due to limited access to high-quality ground truth data and the lack of interpretability and trustworthiness in the obtained solutions. These hurdles have impeded the broad adoption and clinical implementation of deep learning based image reconstruction algorithms. This thesis advocates for the integration of uncertainty quantification as a means to mitigate the drawbacks of deep learning methods applied to computed tomography. We advance deep unrolled optimisation schemes within a Bayesian framework, termed Bayesian deep gradient descent, and propose a novel unsupervised adaptation paradigm for addressing poor generalizability in deep learning solutions. Transitioning to the deep image prior framework is a logical progression. Unsupervised learning avoids the need for high-quality datasets and demonstrates robust performance in out-of-distribution settings. We introduce a novel framework, named the linearised deep image prior, for scalable and calibrated uncertainty quantification. This framework leverages the linearised Laplace approximation while offering valuable insights into unsupervised learning. We benchmark our methods with real-measured computed tomography data, including volumetric image reconstruction despite computational challenges. Finally, we not only provide calibrated uncertainty quantification but also explore its effective utilisation for guiding additional imaging tasks in an information-theoretic manner. We introduce a promising proof-of-concept strategy for intelligent scanning regimes using uncertainty quantification, where the model governs the data collection process in computed tomography, utilising its own understanding of the reconstruction task.

Download activity - last month

Export as XMLJSONCSV

Download activity - last 12 months

Export as JSONXMLCSV

Downloads by country - last 12 months

Export as JSONXMLCSV

Archive Staff Only

View Item