Semi-automated quantification of cartilage volume in knee osteoarthritis: responsiveness and concurrent validity

Abstract

Purpose: Cartilage change on Magnetic Resonance Imaging (MRI) is a potential surrogate endpoint in osteoarthritis (OA) research. Semi-quantitative scoring systems are based on qualitative and subjective assessments that require expensive and time-intensive radiology expertise; ordinal scores also have limited sensitivity to change. Quantitative measurement of cartilage that involves manual segmentation of cartilage is also time-intensive and costly. To overcome these limitations, semi-automated software systems have been developed to measure cartilage - providing a rapid, inexpensive, quantitative and objective method of evaluating disease status and progression. A novel and semi-automated Local-Area Cartilage Segmentation (LACS) software for knee OA has previously been demonstrated to be fast, responsive, and associated with radiographic and pain progression. However, LACS was initially developed to segment medial femur cartilage only. We aimed to update and extend LACS to the lateral femur, medial tibia, and lateral tibia - and evaluate responsiveness to change, as well as correlation with existing cartilage measurements using an established independent manual segmentation method from publicly available Osteoarthritis Initiative (OAI) data. Methods: 115 participants with symptomatic knee OA were selected from the OAI progression sub-cohort, defined by definite osteophytes and frequent knee symptoms. Cartilage volume in fixed weight-bearing areas was measured with LACS on unilateral knee MRIs at the baseline and 24 month visit, paired and blinded to image date, using the sagittal 3D double-echo steady-state (DESS) sequence (3 Tesla, 0.365mm x 0.365mm, 0.7mm slice thickness, repetition time 16.5 msec, echo time 4.7 msec). Briefly, a non-expert reader identified a set of consistent anatomical landmarks for each compartment on the femur and tibia, which were used to generate consistent cartilage measurement areas. Automated edge detection algorithms then outlined cartilage in regions defined by the landmarks (Figure). The expert reader corrected the cartilage margins in areas of misidentified cartilage margins when necessary. Change in cartilage volume was calculated between the baseline and the 24-month visit. Responsiveness was quantified by standardized response mean (SRM = mean (ΔVol)/SD (ΔVol)). Concurrent validity was evaluated by estimating the correlation between LACS volume measurements and Chondrometrics measurements of cartilage thickness based on manual segmentation in the four corresponding sub-regions. Results: Baseline Kellgren-Lawrence grade distribution was KL1 4.3%, KL2 32.2%, KL3 61.7%, and KL4 1.7% in knees of the 115 participants (53% females, mean (SD) age 62.6 (±8.8) years, mean (SD) and BMI 30.1 (±5.3) kg/m2). The Figure shows a rendered 3D-model of segmented cartilage, with red areas indicating selected regions of measurement. The SRM was -0.54 for the medial femoral region and lower in the other sub-regions (Table 1). Correlation with Chondrometrics measures of cartilage thickness ranged between 0.67 and 0.88 (Table 2). The expert reader time was under 5 minutes and total reader time less than 10 minutes total per compartment. Conclusions: We updated the semi-automated LACS software method for knee cartilage segmentation to not only include the medial femur, but also medial tibia, lateral tibia, and lateral femur. The automated steps increase efficiency compared to manual segmentation, and limit the need and cost of an expert reader. We found favorable SRMs for cartilage volume change over two years with default regions of interest. The flexible nature of this software enables further adjustment of sub-regions to potentially improve responsiveness. The software correlated well against established measurements of cartilage thickness based on manual segmentation. Although a comparison between thickness and volume is not ideal, the current results are encouraging. In summary, the updated LACS software represents a powerful tool for fast segmentation of knee cartilage in larger cohorts and could facilitate clinical evaluation of potential structure modifying OA interventions.