Knee osteoarthritis commonly requires joint replacement, substantially reduces quality of life and increases healthcare utilisation and costs. This study aimed to identify whether quantitative measures of articular cartilage structure predict knee replacement, and to establish their utility as outcomes in clinical trials of disease-modifying therapy. A nested case-control study was performed in Osteoarthritis Initiative participants, a multicentre observational cohort of 4796 participants with or at risk of knee osteoarthritis. 127 knees were replaced between baseline and 4 years follow-up, and one control knee per case matched for baseline radiographic disease stage (Kellgren-Lawrence grade; KLG), gender and age. Quantitative cartilage measures were obtained from 3 T magnetic resonance images at the exam before knee replacement, and longitudinal change during the previous 12 months when available (n=110). Cartilage thickness loss in the central and total medial femorotibial compartment (primary and secondary predictor variables) was significantly greater in case than control knees (AUC=0.59/0.58). Differences in cartilage loss were greater at earlier than later radiographic disease stages (p<0.01 for interaction with KLG). Cartilage thickness loss in the central tibia was the most predictive longitudinal measure (AUC=0.64). Denuded bone areas in the medial femur were the most predictive and discriminatory cross-sectional measure between case and control knees (AUC=0.66). This study demonstrates the predictive value of quantitative, MRI-based measures of cartilage for the clinically relevant endpoint of knee replacement, providing support for their utility in clinical trials to evaluate the effectiveness of structure-modifying intervention.