Advanced Imaging Techniques for Knee and Ankle Pathologies: Applications of deep learning and statistical shape modelling

Abstract

This thesis presents innovative applications of deep learning and statistical shape modeling (SSM) in musculoskeletal imaging, with a focus on knee and ankle pathologies, such as osteoarthritis, enthesitis, and impingement. The aim is to advance diagnostic precision and quantification of these conditions while minimizing patient risk, particularly by avoiding radiation exposure. Chapter 1 introduces the challenges in diagnosing musculoskeletal pathologies, underscoring the need for accurate and radiation-free imaging modalities. Conventional imaging techniques like CT and MRI are discussed, emphasizing their limitations in capturing fine anatomical details necessary for diagnosing early-stage diseases. The potential of synthetic CT (sCT), which generates CT-like images from MRI data, is highlighted as a promising alternative. Chapter 2 details the use of SSM for quantifying morphological variations in the talus bone related to ankle impingement. A case-control study compared talus shapes from individuals with and without impingement, identifying key shape modes that distinguish between the groups. This study demonstrates the clinical relevance of SSM in identifying subtle structural differences and potential surgical planning applications. Chapter 3 explores the feasibility of MRI-based sCT in detecting knee osteoarthritis (OA) and compares its diagnostic capabilities to conventional CT. A machine learning algorithm trained on MRI data generated sCT images, achieving comparable diagnostic accuracy for bone-related OA features. The findings support sCT’s potential to reduce radiation exposure in OA imaging while providing comparable diagnostic quality. Chapters 4 and 5 focus on the automatic quantification of small inflammatory pathologies, specifically heel enthesitis and ankle tenosynovitis, using deep learning. These studies address the limitations of manual scoring systems, which are time-consuming and subjective. The developed deep learning models achieved reliable segmentation and quantification of inflammation, paving the way for automated, reproducible assessments of inflammatory arthritis in clinical practice. Chapter 6 discusses the broader clinical implications of these findings, emphasizing how advanced imaging techniques could streamline diagnostic workflows, improve patient outcomes, and reduce healthcare costs. This chapter also addresses future research directions, such as expanding sCT applications to other joints and integrating AI-driven quantification into routine clinical practice. The work culminates in software tools developed to facilitate these imaging techniques, providing accessible, reproducible assessments that support clinical decision-making. This thesis lays the groundwork for advancing musculoskeletal imaging through AI and SSM, offering a vision for radiation-free, accurate diagnostics across diverse clinical contexts.