Knee Cartilage Change on Magnetic Resonance Imaging: Should we Lump or Split Topographical Regions? A Two-Year Study of Data from the Osteoarthritis Initiative

Abstract

Objective:

We challenge the paradigm that a simplistic approach evaluating anatomic regions (e.g., medial femur or tibia) is ideal for assessing articular cartilage loss on magnetic resonance (MR) imaging. We used a data-driven approach to explore whether specific topographical locations of knee cartilage loss may identify novel patterns of cartilage loss over time that current assessment strategies miss.

Design:

We assessed 60 location-specific measures of articular cartilage on a sample of 99 knees with baseline and 24-month MR images from the Osteoarthritis Initiative, selected as a group with a high likelihood to change. We performed factor analyses of the change in these measures in two ways: (1) summing the measures to create one measure for each of the six anatomically regional-based summary (anatomic regions; e.g., medial tibia) and (2) treating each location separately for a total of 60 measures (location-specific measures).

Results:

The first analysis produced three factors accounting for 66% of the variation in the articular cartilage changes that occur over 24 months of follow-up: (1) medial tibiofemoral, (2) medial and lateral patellar, and (3) lateral tibiofemoral. The second produced 20 factors accounting for 75% of the variance in cartilage changes. Twelve factors only involved one anatomic region. Five factors included locations from adjoining regions (defined by the first analysis; e.g., medial tibiofemoral). Three factors included articular cartilage loss from disparate locations.

Conclusions:

Novel patterns of cartilage loss occur within each anatomic region and across these regions, including in disparate regions. The traditional anatomic regional approach is simpler to implement and interpret but may obscure meaningful patterns of change.

Introduction

Articular cartilage damage is a hallmark of knee osteoarthritis, and quantification of this construct is important for researchers assessing disease progression in knee osteoarthritis (Eckstein et al., 2005). The widespread promulgation of magnetic resonance (MR) imaging as a clinical radiologic tool, safe in humans, has led to an interest in quantifying measures of articular cartilage (Duryea et al., 2014; Eckstein, Cicuttini, Raynauld, Waterton, & Peterfy, 2006; Hellio Le Graverand et al., 2009; Hunter et al., 2009). Traditionally, investigators have relied on anatomically regional-based summary statistics to assess cartilage loss (e.g., mean cartilage thickness in the medial tibia)(Eckstein, Ateshian, et al., 2006). However, there may be other meaningful patterns of cartilage loss in areas that are correlated because of biomechanical factors (e.g., arthrokinematics, bone shape).

We hypothesized that examining location-specific measures of cartilage change may yield novel topographic groupings of cartilage loss to identify and study in the context of osteoarthritis-related research. These findings could generate new insights into the natural history of knee osteoarthritis and spark discussion regarding the possibility of considering alternative strategies to analyze articular cartilage. Hence, we performed two exploratory factor analyses, one using six anatomic regions of cartilage change and the other using 60 location-specific measures of cartilage change. The goal was to identify articular cartilage measures that change together over 24 months using these two differing strategies.

Materials & Methods

Study Design

This study was a 2-year longitudinal study of a subcohort within the Osteoarthritis Initiative (OAI).

The Osteoarthritis Initiative (OAI)

The OAI is a multicenter, observational cohort study of people at high risk of developing or already having symptomatic knee osteoarthritis. Participants were recruited from 2004 to 2006 at four clinical sites: University of Maryland (Baltimore, Maryland), Memorial Hospital of Rhode Island/Brown University (Pawtucket, Rhode Island), Ohio State University (Columbus, Ohio), and University of Pittsburgh (Pittsburgh, PA).

This study received ethical approval from each OAI clinical site (Memorial Hospital of Rhode Island Institutional Review Board, The Ohio State University’s Biomedical Sciences Institutional Review Board, University of Pittsburgh Institutional Review Board, and University of Maryland Baltimore – Institutional Review Board), and the OAI coordinating center (Committee on Human Research at University of California, San Francisco). The OAI adhered to the ethical standards laid down in the 1964 Declaration of Helsinki and its later amendments. All persons gave their informed consent before their inclusion in the study.

Study Sample

The OAI investigators created a Core Image Assessment sample from which we selected participants for our parent study (J. B. Driban et al., 2022). We included one knee from each participant with symptomatic knee osteoarthritis at the OAI baseline visit (frequent knee pain and Kellgren-Lawrence (KL) grade 2 or 3) and complete baseline and 24-month data (clinical, radiographic, and MR outcomes). The original selection of this group used KL grades that were read locally at each clinical site. However, the KL grades reported in this manuscript relied on updated central radiographic readings, explaining why a few people had a KL grade = 0, 1, or 4. The reliability of these central readings (read-reread) was good. Within each KL grade, we randomly sampled participants with and without radiographic progression (any increase in KL grade over 24 months). The final sample size was 99 participants, which we described in a prior publication (J. B. Driban et al., 2022).

Radiographic Assessments

Participants had annual bilateral weight-bearing, fixed-flexion posterior-anterior knee radiographs. Central readers reported medial and lateral joint space narrowing grades and KL grades (0 to 4; intra-rater agreement using weighted kappa ranged from 0.70 to 0.88). Data and protocols are publicly available at https://data-archive.nimh.nih.gov/oai (kxr_sq_bu00 (version 0.8) and kxr_sq_bu03 (version 3.7)).

Magnetic Resonance Imaging

All participants of this study had MR imaging on the same knee at the OAI baseline and 24-month follow-up visits. Three-dimensional Double Echo Steady State (3D DESS) sequences had to be available to be included in this sample. The 3D DESS sequences were acquired using the standard parameters: field of view = 140mm, slice thickness = 0.7mm, skip = 0mm, flip angle = 25 degrees, echo time = 4.7ms, recovery time = 16.3ms, 307 × 384 matrix, x resolution = 0.365mm, y resolution = 0.456mm, total slice number = 160. All OAI MR images were obtained using identical scanners at each clinical site. The study protocol and images are publicly available at https://nda.nih.gov/oai.

Measurement of Cartilage Damage Index (CDI) on MR Images

We used the CDI, a validated method (Zhang et al., 2016; Zhang et al., 2014; Zhang et al., 2015), to quantify cartilage damage on the 3D DESS MR images from the baseline and 24-month visits. The CDI has previously been successfully used in clinical trials and observational studies to evaluate for differences between groups (Harkey et al., 2020; Lo et al., 2022; McAlindon et al., 2017). These locations (Figure 1) are robust, and we have shown that among knees with early (pre-radiographic osteoarthritis), our cartilage measurements are prognostic of an accelerated onset and progression of knee osteoarthritis (Harkey et al., 2019; Harkey et al., 2020) and future knee replacement (J. Driban et al., 2019). We used semi-automated software (Zhang et al., 2016; Zhang et al., 2014; Zhang et al., 2015) to delineate the bone-cartilage boundary on automatically selected images. We also used the software to automatically identify the predefined informative locations to measure cartilage thickness in the six anatomic regions (medial and lateral tibia, medial and lateral femur, and medial and lateral patella). There were nine locations in tibiofemoral regions, 12 locations in patellar regions; 60 locations total. We measured the CDI by calculating the products of cartilage thickness, cartilage length (the length of the interface between the cartilage and subchondral bone on each sagittal slice), and voxel size from each informative location. Anatomic regional measures of CDI represent the sum of all CDI measures within that region. Our reader (MZ) had good reproducibility for the CDI with inter-tester (intra-class correlation coefficient [ICC]_2,1) reliability of 0.86 - 0.95 and intra-tester (ICC_3,1) reliability of 0.94 - 0.99. The reader was not blinded to the chronological order of the images.

Figure 1. Cartilage Damage Index (CDI) Measurement Locations.

In brief, locations 1-9 are in the medial femur, 10-18 lateral femur, 19-27 medial tibia, 28-36 lateral tibia, 37-48 medial patella, and 49-60 lateral patella. In the tibiofemoral compartment, the lowest number in each anatomic region is the most anterior-medial location and the highest is the most posterior-lateral location. In the patellar regions, the lowest number in each anatomic region is the most medial-inferior location and the highest is the most lateral-superior location. Please return to this figure to understand the location-specific factors are identified in Table 3.

Statistical Analysis Plan

We performed factor analyses, a multivariate statistical method that facilitates grouping of the smallest number of theoretic constructs (otherwise known as “factors”) that can most efficiently explain the co-variation observed within a set of measured variables (Watkins, 2018). In this type of analysis, a factor is an unobservable variable that impacts greater than one observed measure and explains associations among these observed measures. The observed measures are associated because they share some underlying commonality.

For the first factor analysis, we used change in CDI measures from baseline to 24 months, using the six anatomic regions. As part of the factor analysis, an orthogonal rotation was applied. Factors with an eigenvalue > 1 were retained. Anatomic regions with factor loadings greater than the absolute value of 0.4 were considered to contribute to a specific latent variable (factor). We reviewed each anatomic region that contributed to each factor to identify the respective underlying construct.

For the second factor analysis, we included each of the 60 informative location changes in CDI measures, following similar methods as described above. When reviewing the location-specific changes in CDI that contributed to each latent factor, we used the results from the first factor analysis to define “adjoining” and “disparate” regions for this second factor analysis. Adjoining regions represented anatomic regions that grouped together in a factor in the first factor analysis. In contrast, disparate regions were anatomic regions that did not group together in the first factor analysis.

We used SAS 9.4 for all analyses.

Results

Our sample included 99 knees. The knee frequencies for each KL grade were as follows: 1 (~1%) with KL grade 0, 9 (~9%) with KL grade 1, 40 (~40%) with KL grade 2, 47 (~47%) with KL grade 3, and 2 (~2%) with KL grade 4. 66% had medial joint space narrowing grade > 0 and 25% had lateral joint space narrowing grade > 0. Nearly a third (30%) of knees had an increase of at least one KL grade over 24 months. Our sample was primarily female (60%), with an average age of 61 (9) years and a mean body mass index of 30.2 (4.6) kg/m² (Table 1).

Table 1:

Descriptive Characteristics of the Sample of 99 Participants.

Descriptive Characteristics
Age, mean (SD), y	61.5 (8.7)
Female, n (%)	59 (60%)
Body Mass Index (kg/m2), mean (SD)	30.2 (4.6)
Kellgren-Lawrence Grade, Baseline, n (%)
0	1 (1%)
1	9 (9%)
2	40 (40%)
3	47 (48%)
4	2 (2%)
Medial Joint Space Narrowing Grade, Baseline, n (%)
0	34 (34%)
1	28 (28%)
2	36 (36%)
3	1 (1%)
Lateral Joint Space Narrowing Grade, Baseline, n (%)
0	74 (75%)
1	12 (12%)
2	12 (12%)
3	1 (1%)
Progression * , n (%)	29 (29%)
Non Progression * , n (%)	70 (71 %)

Note:

^*Progression was defined as increasing at least 1 KL grade from baseline to 24-month follow-up visit.

Factor Analysis by Anatomic Region

The factor analysis by anatomic region, including the six anatomic regions, produced three latent factors, accounting for 66% of the variance in CDI change; the higher the percentage of variance in CDI change explained by the factors included, the better the performance of the groupings by selected factors (Table 2). The three factors that emerged were (1) medial tibiofemoral, (2) lateral tibiofemoral, and (3) medial and lateral patella.

Table 2.

Factor Analysis by Anatomic Region

Anatomic Region	Factor 1 (Medial Tibiofemoral Compartment)	Factor 2 (Patellar Regions)	Factor 3 (Lateral Tibiofemoral Compartment)
Medial Femur	0.795	0.128	−0.197
Lateral Femur	−0.039	−0.073	0.778
Medial Tibia	0.828	−0.073	0.039
Lateral Tibia	−0.085	0.112	0.719
Medial Patella	0.089	0.787	0.133
Lateral Patella	−0.049	0.799	−0.085

Note: Factor loadings greater than the absolute value of 0.4 were considered to contribute to a specific latent variable and are in bold

Factor Analysis by Specific Locations

The second factor analysis, which included 60 location-specific measures, produced 20 latent factors (Table 3), which accounted for 75% of the variance in CDI changes. The first three factors accounted for 23% of the total variation and only involved locations in the patella. Twelve factors only involved one anatomic region. Five factors included locations from adjoining regions (e.g., factor 7 – medial tibia with medial femur). Three factors included locations from disparate locations (e.g., factor 12 – medial patella with lateral femur).

Table 3.

Factor Analysis by Location Specific Measures

Factor	Anatomic Regions Included	CDI Locations Included (see Figure 1)	One Regiona	Adjoining Regionsb	Disparate Regionsc
Factor 1	Medial Patella Lateral Patella	42, 43, 44, 47, 48 59		x
Factor 2	Lateral Patella	51, 52, 56, 60	x
Factor 3	Lateral Patella	49, 50, 53, 54, 55, 57, 58	x
Factor 4	Medial Tibia	19, 20, 22, 23, 25	x
Factor 5	Medial Patella Lateral Patella	41, 42, 45, 46 57		x
Factor 6	Lateral Femur Lateral Tibia	11, 16, 17 35		x
Factor 7	Medial Femur Medial Tibia	1, 2, 4, 9 25		x
Factor 8	Lateral Tibia	31, 32, 33	x
Factor 9	Lateral Tibia	28, 29	x
Factor 10	Medial Femur Medial Tibia	3 21, 24		x
Factor 11	Lateral Femur	18	x
Factor 12	Lateral Femur Medial Patella	15 37, 38			x
Factor 13	Lateral Femur	12	x
Factor 14	Medial Tibia Medial Patella	27 39			x
Factor 15	Medial Femur Lateral Femur	5 10			x
Factor 16	Lateral Femur	13	x
Factor 17	Medial Patella	40	x
Factor 18	Medial Tibia	26	x
Factor 19	Lateral Femur	14	x
Factor 20	Medial Femur	8	x

Note: CDI = cartilage damage index.

^aOne region: one of 6 possible anatomic regions: medial or lateral patella, femur, or tibia.

^bAdjoining Regions: anatomic regions that clustered together in the factor analysis only including 6 anatomic regions (i.e., medial tibiofemoral, lateral tibiofemoral, and medial and lateral patella)

^cDisparate Regions: more than one anatomic region not in an adjoining region.

Discussion

In this study, the traditional anatomically regional-based approach yielded three factors that explained a substantial amount of the variation of cartilage change (66%). This factor analysis yielded factors that we expected: the respective medial and lateral tibiofemoral compartments and the patellar regions (the medial and lateral patella). These three factors reaffirm that when studying anatomic regions of articular cartilage, there’s merit to lumping them by compartment (e.g., medial tibiofemoral). This approach is appealing because it is simple to implement and interpret. However, our second factor analysis, which explained more of the variation in cartilage change, identified novel patterns of cartilage loss within each anatomic region and across these regions, including in disparate regions.

On the one hand, the traditional approach (that uses the concept of lumping) may obscure meaningful patterns of change that warrant further attention but is simple to implement. On the other hand, the novel approach of considering multiple anatomic cartilage measures separate (that uses the concept of splitting) requires more effort to implement, but may also allow for the detection of patterns of cartilage change that would otherwise not be visible using traditional methods. These findings suggest that there might be differential benefits and downsides to using different strategies for studying articular cartilage in osteoarthritis. No uniform answer applies to all studies. Ultimately, researchers need to decide which approach is best for their project. This decision depends on their research questions and whether those questions would benefit from the traditional approach, which is simpler and generalizable, or the location-specific approach, which offers a greater spatial understanding of cartilage change, which may be particularly helpful in biomechanical studies.

When looking at all the locations individually, we identified that 12 out of 20 factors involved one specific point or a discrete area within an anatomic region (e.g., Factor 2: superior lateral patella). Additionally, there were five groupings where change occurred in adjoining regions, but this was not the most common pattern. Finally, we observed unexpected groupings of articular cartilage that changed together in disparate regions within the knee. These findings supported our hypothesis that examining location-specific measures of cartilage change may yield novel topographic groupings of cartilage changes.

Other investigators have previously suggested that there may be merit in focusing on anatomical-based subregions as those may be more responsive to change. For example, cartilage reduction was most common in the central and external subregions of the medial weight-bearing femur in a different substudy to the OAI (Buck, Wirth, Dreher, Nevitt, & Eckstein, 2013). LeGraverand et al. found participants with a higher KL grade had significantly thinner cartilage than those with lower KL grades, especially in the medial weight-bearing femur, in the external subregion of the medial tibia, and in the internal subregion of the lateral tibia (Hellio Le Graverand et al., 2009). Eckstein et al. have suggested that the central medial femur is one of the most responsive regions to change (Eckstein, Ateshian, et al., 2006; Eckstein et al., 2005). While prior studies highlight that regional summary statistics may dilute the responsiveness of important subregional changes, our study adds that there may be other ways to characterize unique subregions that have not been previously described. Additionally, these prior studies focused on specific regions using a priori hypotheses as opposed to what we have done in our study, where our findings were data driven. We allowed the data to show us the patterns of change that were present.

Notably, the findings from this study highlight that while there has traditionally been a primary focus on the tibiofemoral cartilage in clinical trials and observational studies of knee OA (Collins et al., 2022; McAlindon et al., 2017; Reichenbach et al., 2010), other regions, such as the patella should be considered when discriminating rates of change in cartilage. This is supported by the fact that there is evidence from other studies that knees that develop widespread structural damage or osteoarthritis are more likely to start with it isolated to the patellofemoral joint (Lankhorst et al., 2017; MacKay et al., 2021; Stefanik et al., 2016). Given these findings and our results that medial and lateral patellar regions accounted for significant variance in cartilage change, these areas should be focused on in future analyses.

Although our study highlights the importance of considering novel topographical cartilage groupings as a complementary method to the traditional regional summary of cartilage change, there are some limitations. Notably, we relied on 24-month change in the CDI among 99 knees selected for the presence of radiographic tibiofemoral osteoarthritis without imaging the patellofemoral joint. Hence, the generalizability and reproducibility of the identified groupings may be affected if we 1) selected knees based on tibiofemoral and patellofemoral osteoarthritis, 2) accounted for osteoarthritis phenotypes that may influence correlations between cartilage loss in different regions (e.g., history of injury, static alignment), 3) relied on a different follow-up duration than is common in clinical trials, 4) and included articular regions less susceptible to full-thickness cartilage defects, which are omitted in the CDI. Furthermore, we have not tested the validity or sensitivity to change of these new groupings. However, this exploratory analysis confirmed that there might be novel topographical latent factors that future studies should explore in larger studies designed to validate these measures, determine their sensitivity to change, and explore various factors that could influence topographical cartilage groupings.

Novel patterns of cartilage loss occur within each anatomic region and across these regions, including in disparate regions. For the last 25 years, researchers have generally taken an average of the overall cartilage loss and assumed it represented one construct. These findings suggest that this approach is not always best since sometimes all the cartilage does not represent just one construct. Sometimes, other locations distant from the original location group together. The traditional anatomically-based region approach is simpler to implement and interpret but may obscure meaningful patterns of change.

Conclusion

In conclusion, our findings support the idea that the traditional anatomic region approach of quantifying cartilage change is a reasonable method to obtain meaningful data on disease progression. However, the traditional approach may obscure meaningful patterns of change within a region and across regions, lending to the idea that a study of articular cartilage that is focused on location-specific measures can offer opportunities to explore novel patterns of cartilage change not otherwise observed. Future studies can complement our exploratory factor analysis findings by exploring the validity or sensitivity to change of novel patterns of cartilage change. By providing a higher level of detail about specific topographical points and regions of change, our results can be useful in comparing groups (e.g., treatment versus control groups in disease-modifying OA drug studies) and evaluating clinical outcomes. The selection of approach regarding how to measure articular cartilage, whether using the anatomically regional-based approach or the location-specific measure approach, should depend on the hypothesis being tested.

Data Availability Statement:

The datasets generated during and/or analyzed during the current study are available from the corresponding author upon reasonable request. Datasets from the main Osteoarthritis Initiative are publicly available on the website: https://nda.nih.gov/oai/