Pascale, Michele; (2025) Unsupervised Learning for Generalised Super-Resolution of 3D Anisotropic Medical Images via Domain Transfer. Masters thesis (M.Phil), UCL (University College London).

Preview

Text thesis.pdf - Accepted Version Download (22MB) | Preview

Abstract

Three-dimensional (3D) imaging plays a crucial role in various medical applications, particularly in fields such as radiology and neurology, where detailed anatomical insights are crucial. However, in many clinical settings, anisotropic 3D volumes—characterized by uneven spatial resolution across different axes—are commonly acquired to balance scan quality and acquisition time. These anisotropic datasets often result from thick slices with low spatial resolution, which, while reducing scan time, can pose challenges for accurate analysis and interpretation. Deep learning (DL) offers a solution to recover high-resolution features through super-resolution reconstruction (SRR). Unfortunately, paired training data is unavailable in many 3D medical applications and therefore a novel unpaired approach is proposed; CLADE (Cycle Loss Augmented Degradation Enhancement). CLADE uses a modified CycleGAN-based architecture with a cycle-consistent gradient mapping loss and weight demodulation process. This approach is trained in an unsupervised fashion to learn SRR of the low-resolution dimension, from disjoint patches of the high-resolution plane within the anisotropic 3D volume data itself via resolution domain transfer. The feasibility of CLADE in abdominal Magnetic Resonance Imaging (MRI) and abdominal Computed Tomography (CT) imaging is demonstrated, with improvements in CLADE image quality over low-resolution volumes, conventional Cycle-GAN and state-of-the-art self-supervised SRR; SMORE (Synthetic Multi-Orientation Resolution Enhancement).

Download activity - last month

Export as CSVXMLJSON

Download activity - last 12 months

Export as JSONXMLCSV

Downloads by country - last 12 months

Export as CSVXMLJSON

Archive Staff Only

View Item