Automatic framework for evaluating osteoarthritic cartilage severity: high-resolution cartilage thickness mapping and scoring

ObjectivesTo develop and validate an automatic, scalable framework for assessing the femoro-tibial osteoarthritic cartilage severity using high-resolution cartilage thickness maps (CTh-Maps) and a cartilage thickness scoring system (CTh-Score). MethodsThe Osteoarthritis Initiative (OAI) cohort of 4796 subjects was analyzed. A 3D-UNet was trained to segment femoro-tibial bones and cartilages using MRI from baseline, 1-, 2-, 3-, 4-, 6- and 8-year follow-ups. CTh-Maps were created for each knee. A ResNet model trained on CTh-Maps assigned a CTh-Score ranging from 0 (healthy cartilage) to 100 (end-stage OA). The reproducibility of the CTh-Score was evaluated in a test/retest setup. Its validity was assessed by examining the correlation with expert evaluations of cartilage loss (MOAKS grading) and association to OA severity (KL grade) in both OAI and external dataset. The CTh-Score sensitivity to OA structural progression was examined. ResultsThe framework generated CTh-Maps for the entire OAI, forming the "OAI CTh-Maps" dataset. Both CTh-Maps and CTh-Score showed excellent reproducibility (ICC>0.98). The CTh-Score demonstrated strong correlations (r=0.81) with expert assessments of cartilage loss and strong associations to OA severity, including in the external dataset. The CTh-Score either increased or remained stable for almost all subjects at 8-year follow-up. The CTh-Score showed great sensitivity to change, significantly increasing between each timepoint, up to 6 years prior to KL progression. ConclusionsCTh-Maps and CTh-Score represent a novel approach to analyze cartilage at imaging. Their scalability, reproducibility and sensitivity to osteoarthritic cartilage severity provide significant opportunities for earlier OA detection, better disease monitoring, and therapeutic window identification.