Radiomic signatures from plain radiographs predict knee osteoarthritis progression : from single- to multi-view analysis

Abstract

Background: Knee osteoarthritis (OA) is a prevalent joint disorder characterised by progressive structural degeneration and heterogeneous disease trajectories. It involves both the tibiofemoral (TF) and patellofemoral (PF) compartments, often with uneven severity, creating challenges for consistent diagnosis and personalised management. Traditional radiographic assessments, such as the Kellgren-Lawrence (KL) grading system, show limited sensitivity, particularly for compartment-specific or early-stage OA. The KL system also suffers from substantial interobserver variability, weak correlation with clinical outcomes, and a non-linear categorical scale that restricts quantitative interpretation. Given the need for scalable, cost-effective, and objective tools, radiomics—the quantitative analysis of imaging features—offers a promising avenue for personalised OA risk prediction. This thesis develops and validates radiomics-based biomarkers for assessing knee OA severity and progression, focusing on compartment-specific modelling and multi-view feature integration. Methods: Data were derived from multiple large, multicentre knee OA cohorts, including the Multicenter Osteoarthritis Study (MOST), the Osteoarthritis Initiative (OAI), the Hong Kong EHR-derived Knee OA Cohort, the Meniscal Tear and Osteoarthritis Risk (MenTOR) Cohort, and the Knee Injury Cohort at the Kennedy (KICK). All provided standardised radiographs and longitudinal clinical outcomes. (i) Radiomic features were first extracted from the PF joint using semi-automated segmentation and advanced image-processing techniques to build a predictive model for future knee replacement. (ii) To enhance cross-cohort generalisability, a domain adaptation strategy was implemented to harmonise differences in imaging protocols and population characteristics. (iii) A deep-learning-based radiomics framework was then developed for the TF joint, employing convolutional neural networks to capture complex structural patterns beyond those represented by KL grading. (iv) Finally, a multi-view learning approach combined PF and TF radiomics features to evaluate whether integrated compartmental information improved prediction of OA progression. Results: Four principal findings were obtained. (i) The PF radiomics score demonstrated additive predictive value to the KL grade, with their combination achieving an area under the receiver operating characteristic curve (AUC) of 0.87 versus 0.84 for KL alone (p < 0.001). (ii) The domain-adapted PF model showed strong external generalisability, yielding AUCs of 0.73 (US), 0.70 (Hong Kong), and 0.64 (UK), surpassing alternative models. (iii) The deep-learning TF radiomics framework improved assessment of OA severity and progression, reaching a concordance index (C-index) of 0.85 and an AUC of 0.89, indicating enhanced sensitivity to structural change. (iv) The integrated PF–TF radiomics model further improved predictive accuracy, achieving a C-index of 0.91 and an AUC of 0.93, underscoring the importance of comprehensive, compartment-aware radiographic analysis. Conclusion: Radiomics analysis of the PF and TF joints offers a powerful, cost-effective, and scalable imaging-based tool for personalised knee OA risk prediction. Through advanced feature extraction, deep learning, and domain adaptation, the proposed models show robust generalisability and clinical relevance. The integration of multi-compartmental radiomic features further enhances predictive precision, enabling comprehensive assessment of disease status. Overall, this thesis refines current radiographic evaluation methods and establishes a methodological framework that may support future precision-medicine approaches, promoting earlier detection and more individualised management of knee OA.