Clinical Validation of Artificial Intelligence (AI)-based Cartilage Segmentation Predicting Knee Replacement

For cartilage morphology to serve as a scalable endpoint in clinical trials, analyses should be performed automatically without human interaction. To clinically validate artificial intelligence (AI)-based analysis, we studied cartilage loss from MRI prior to knee replacement.

Knee replacement cases between 36 and 60 months' follow-up in the Osteoarthritis Initiative were matched with controls by age, sex and radiographic status. Cartilage thickness was determined by sagittal DESS and coronal FLASH MRI. Segmentation was performed with automated (AI-based: DESS and FLASH) and manual (DESS) methods. Two-year change prior to knee replacement was compared using Cohen's D, the medial compartment defined as the primary analytic endpoint: RESULTS: One-hundred-thirty-two case and matched control knees (113 participants; 55% women; age 64.4±8.6 [mean ± SD] years) had DESS, and 35 had FLASH in both the case and control knee. In case knees (DESS), automated segmentation revealed an MFTC cartilage thickness change of -263μm (SRM -0.63), and manual analysis -259μm (SRM -0.64). In controls, the rate of change was substantially lower by either method, with the effect size similar for automated vs. manual segmentation (Cohen's D -0.40 vs. -0.35). With automated FLASH, the Cohen's D was -0.59 vs. -0.30 by the manual method.

AI-based cartilage segmentation performed at least as well as manual expert analysis in separating knees with subsequent knee replacement from controls. This was consistent across MRI acquisitions with different orientations and contrasts. The findings support automated MRI-based cartilage morphometry as a scalable imaging biomarker associated with a clinically meaningful outcome.