Investigating the relationship between radiographic joint space width loss and deep learning-derived magnetic resonance imaging-based cartilage thickness loss in the medial weight-bearing region of the tibiofemoral joint

Abstract

Objective

To investigate the relationship between measures of radiographic joint space width (JSW) loss and magnetic resonance imaging (MRI)-based cartilage thickness loss in the medial weight-bearing region of the tibiofemoral joint over 12–24 months. To stratify this relationship by clinically meaningful subgroups (sex and pain status).

Design

We analyzed a subset of knees (n ​= ​256) from the Osteoarthritis Initiative (OAI) likely in early stage OA based on joint space narrowing (JSN) measurements. Natural logarithm transformation was used to approximate near normal distributions for JSW loss. Pearson Correlation coefficients described the relationship between ln-transformed JSW loss and several versions of deep learning-derived MRI-based cartilage thickness loss parameters (minimum, maximum, and mean) in subregions of the femoral condyle, tibial plateau, and combined femoral and tibial regions. Linear mixed-effects models evaluated the associations between the ln-transformed radiographic and MRI-derived measures including potential confounders.

Results

We found weak correlations between ln-transformed JSW loss and MRI-based cartilage thickness ranging from R ​= ​−0.13 (p ​= ​0.20) to R ​= ​0.26 (p ​< ​0.01). Correlations were higher (still poor) among females compared to males and painful compared to non-painful knees. Model results showed weak associations for nearly all MRI-based measures, ranging from no association to β (95% CI) ​= ​0.25 (0.11, 0.39). Associations were higher among females compared to males and minimal differences between painful and non-painful knees.

Conclusions

Despite its recommended use in disease-modifying OA drug clinical trials, results suggest that JSW loss is an ineffective proxy measure of cartilage thickness loss over 12–24 months and within a localized region of the tibiofemoral joint.

1. Introduction

Knee osteoarthritis (OA) progression is typically determined through radiographic measures, which offer a time-efficient, inexpensive, and widely accessible tool to visualize the bony aspects of the joint [[1], [2], [3]]. Radiography has, however, faced criticism for its inherent limitations in being able to detect longitudinal changes to the joint throughout the disease process and to the soft tissue aspects of the joint [3,4].

Still, radiographs are the imaging modality of choice in large clinical trials investigating potential disease-modifying OA drugs [3,5]. Currently, no DMOAD has been approved for use to slow or prevent OA progression. The US Food and Drug Administration recommends that DMOAD clinical trials use quantitative radiographic joint space width (JSW), a proxy measure of cartilage thickness, to determine participant eligibility criteria and study endpoints [3,6,7]. It has been suggested that the limitations of these radiographically-defined criteria have contributed to the high failure rates of DMOAD trials and the lack of any DMOAD approved for use in the US [3,8,9].

MRI provides a 3-dimensional visualization of the joint and allows for the assessment of soft tissues, providing a more complete basis for measuring OA disease progression compared to radiography [3,4]. MRI provides direct visualization of cartilage thickness with greater reproducibility compared to radiographic JSW, which is affected by the meniscus and by positioning errors [8,10,11]. Results from previous studies indicate that JSW is only representative of cartilage thickness [7,12,13]. However, how the relationship between JSW and MRI-based measures of cartilage thickness changes over time and within localized regions is less well understood [14]. Using data from the Osteoarthritis Initiative (OAI), this study seeks to quantify associations between JSW loss and MRI-derived cartilage thickness loss within the medial weight-bearing region of the tibiofemoral joint. Given that women are more likely to experience OA than men [15] and the relationship between pain and disease progression is non-linear [16], we a priori investigated the relationship between JSW loss and MRI-based cartilage thickness stratified by biological sex and pain symptom status.

2. Methods

2.1. Study design

This study analyzed data from the National Institutes of Health (NIH)-sponsored OAI. Detailed recruitment information for the OAI has previously been published [17]. Briefly, the OAI includes participants aged 45–79 with or at risk for OA. The OAI was approved by institutional review boards at each enrollment site, with each participant providing informed consent [17].

For this analysis, we selected knees that transitioned from a semi-quantitative joint space narrowing (JSN) score of 0 to ≥1 between any two consecutive timepoints across the first 96 months of OAI follow-up (n ​= ​256). We used this restriction because individuals in the OAI were enrolled at all stages of knee OA, meaning that OAI baseline did not necessarily indicate the beginning of disease progression for all participants. Thus, this inclusion criterion allowed us to re-align OAI participants’ disease progression timelines to clinically meaningful starting points (Fig. 1). The new baseline timepoint (denoted as T₀) used in this analysis is characterized as the last OAI follow-up where a knee received a JSN score of 0. The new endpoint timepoint (denoted as T₁) is characterized as the first OAI follow-up where a knee received a JSN score of at least 1. T₀ and T₁ were 12 or 24 months apart.

Fig. 1.

Visualization comparing the original OAI study timeline and the new re-aligned timeline used in this analysis. T₀ denotes the last timepoint where a knee received a JSN score of 0. T₁ denotes the first timepoint where a knee received a JSN score of at least 1.

2.2. Participant sociodemographic and clinical characteristics

Sociodemographic characteristics included age, sex, body mass index (BMI), knee pain determined through the Knee Injury and Osteoarthritis Score (KOOS) pain subscale, and history of previous knee injury. The KOOS is a self-administered questionnaire with five subscales, including the pain subscale used here, that has been shown to be a reliable measure of symptoms and function in knee OA patients [18,19]. The pain subscale is comprised of nine items and is scored on a 0–100 scale [19]. History of previous knee injury, a significant risk factor for OA development, was based on a self-report “yes/no” question on whether the participant had experienced knee injury that limited their ability to walk for at least 2 days [20]. At OAI baseline, individuals were asked if they ever experienced this type of injury, while at each follow-up timepoint they were asked about whether such an injury occurred within the past 12 months. These covariates were decided on based on previous literature.

2.3. Radiographs

Postero-anterior fixed-flexion weight-bearing bilateral knee radiographs at baseline and follow-up were independently evaluated for semi-quantitative JSN by two trained expert readers (blinded to image timepoint) [17]. JSN is assessed on a 0–3 scale, where 0 indicates a normal joint, 1 suggests mild narrowing, and 2–3 constitutes moderate to severe narrowing [4]. Inconsistencies in measurements were reviewed together by both readers, with the more senior reader determining the final value if needed.

In contrast to JSN, radiographic JSW is a semi-automated and fully quantitative gauge of JSN measured in millimeters (mm). In the OAI, JSW was measured at each follow-up. For this study, we chose to use fixed JSW measurements from the x ​= ​0.250 location in the medial joint compartment, previously reported as the most reliable and responsive measure of OA progression in the medial knee joint [21]. JSW loss (ΔJSW) was calculated as the difference in JSW between T₀ and T₁.

$$\Delta JSW={JSW}_{T0}-{JSW}_{T1}$$

2.4. MRI

Knee MRIs were conducted using 3.0 ​T ​S Trio MR scanners [17]. Cartilage thickness data were obtained through novel deep learning methods [22]. The deep learning process extracted 2-dimensional (2D) cartilage thickness maps from 3-dimensional (3D) knee MRI. Each knee received an annual 2D map, with each pixel on the map representing a millimeter value of cartilage thickness. For this analysis, we used MRI-based cartilage thickness measurements from the medial weight-bearing region of the tibiofemoral joint, which has been shown to experience greater loss of cartilage compared to other regions [23]. We created eight different cartilage thickness loss measures using this MRI-derived data from the medial weight-bearing region of the tibiofemoral joint: three measures used cartilage thickness data from femoral condyle (corresponding to Region 5 in Fig. 2) alone, two used cartilage thickness data the tibial plateau (corresponding to Regions 11 and 12 in Fig. 2), and three used cartilage thickness data from the combined femoral condyle and tibial plateau (corresponding to Regions 5, 11, and 12 in Fig. 2).

Fig. 2.

A) Visualization of the femoral condyle. The weight-bearing region is denoted as 5. B) Visualization of the tibial plateau. The corresponding weight-bearing regions are denoted as 11 and 12. (Image credit: Marc Niethammer).

Minimum cartilage loss measures were calculated as the difference between the minimum cartilage thickness within the specified region between T₀ and T₁. ΔMinFemCTh and ΔMinFemTibCTh represent the minimum cartilage thickness loss within the femoral condyle and combined femoral condyle and tibial plateau medial weight-bearing region, respectively. We calculated the minimum cartilage thickness within the tibial plateau but did not use this measure in the analysis as each observation's value was 0 ​mm. Mean cartilage loss measures were calculated as the difference between the mean cartilage thickness within the specified region between T₀ and T₁. ΔMeanFemCTh, ΔMeanTibCTh, and ΔMeanFemTibCTh represent the mean cartilage thickness loss within the femoral, tibial, and combined femoral and tibial weight-bearing region, respectively. Maximum cartilage loss measures were calculated as the difference between the maximum cartilage thickness within the specified region at T₀ and the minimum cartilage thickness within the specified region at T₁. ΔMaxFemCTh, ΔMaxTibCTh, and ΔMaxFemTibCTh represent the maximum cartilage thickness loss within the femoral, tibial, and combined femoral and tibial weight-bearing region, respectively.

Each measure was calculated as follows:

$$\Delta \mathit{Min}\text{Fem}CTh={minFemCTh}_{T0}-{minFemCTh}_{T1}$$

$$\Delta Mean\text{Fem}CTh={meanFemCTh}_{T0}-{meanFemCTh}_{T1}$$

$$\Delta \mathit{Max}\text{Fem}CTh={maxFemCTh}_{T0}-{minFemCTh}_{T1}$$

$$\Delta Mean\text{Tib}CTh={\text{mean}[TibCTh11}_{T0},{TibCTh12}_{T0}]-{\text{mean}[TibCTh11}_{T1},{TibCTh12}_{T1}]$$

$$\Delta \mathit{Max}\text{Tib}CTh={\max [TibCTh11}_{T0},{TibCTh12}_{T0}]-{\min [TibCTh11}_{T1},{TibCTh12}_{T1}]$$

$$\Delta \mathit{Min}\text{FemTib}CTh=\left({minFemCTh}_{T0}+\min \left[{TibCTh11}_{T0},{TibCTh12}_{T0}\right]\right)-\left({minFemCTh}_{T1}+\min \left[{TibCTh11}_{T1},{TibCTh12}_{T1}\right]\right)$$

$$\Delta Mean\text{FemTib}CTh=\left({meanFemCTh}_{T0}+\text{mean}\left[{TibCTh11}_{T0},{TibCTh12}_{T0}\right]\right)-\left({meanFemCTh}_{T1}+\text{mean}\left[{TibCTh11}_{T1},{TibCTh12}_{T1}\right]\right)$$

$$\Delta \mathit{Max}\text{FemTib}CTh=\left({maxFemCTh}_{T0}+\max \left[{TibCTh11}_{T0},{TibCTh12}_{T0}\right]\right)-\left({minFemCTh}_{T1}+\min \left[{TIbCTh11}_{T1},{TibCTh12}_{T1}\right]\right)$$

Note: TibCTh11 represents tibial cartilage thickness in Region 11, while TibCTh12 represents tibial cartilage thickness in Region 12. Regions 11 and 12 represent the medial tibial plateau, so we used the minimum, mean, and max thickness between the two regions for tibial cartilage thickness values in calculations.

2.5. Statistical analysis

Sociodemographic and participant characteristics were analyzed at the T₀ timepoint and compared between males and females. The normality of data was determined by visual inspection (Supplemental Materials). Natural logarithm transformation was used to produce near-normal distribution for ΔJSW (lnΔJSW) for analytical purposes. Means and standard deviations or frequencies were calculated for continuous and categorical variables, respectively.

For descriptive purposes, Pearson correlation coefficients were calculated between lnΔJSW and each cartilage thickness measure. Next, we used linear mixed-effects models to assess the relationship between the two measures. These models were chosen because they are able to account for within-subject correlations (i.e., the inclusion of two knees belonging to the same individual). The models, conducted with random intercepts per participant, were used to understand the relationship between lnΔJSW (dependent variable) and MRI-derived cartilage thickness loss (independent variable) between T₀ and T₁. Beta coefficients (β) from models were used as a measure of the strength of association between lnΔJSW and MRI-based cartilage thickness loss. β values can be interpreted as the mm change in lnΔJSW associated with a 1 ​mm change in the MRI-based cartilage thickness version (after adjusting for the noted covariates). We used models with unstructured covariance, a decision guided by our understanding of the data and comparing Bayesian Information Criteria and Akaike’s Information Criteria values with other potential structures. We performed multiple imputation via the SAS MI and MIANALYZE procedures, specifying 20 imputations. A Markov chain Monte Carlo method was used, with a separate chain for each imputation.

A priori, we conducted analyses in the clinically meaningful subgroups of sex (male/female) and pain symptom status (painful/non-painful knees). Painful and non-painful knee groups were determined through dichotomizing the KOOS pain subscale, where scores ≥72.2 indicated non-painful knees and scores <72.2 indicated painful knees [24]. All analyses were conducted using SAS Version 9.4 (SAS Institute, Cary NC).

3. Results

In the OAI, 256 knees met our described inclusion criterion (Table 1). This cohort included 212 individuals with 1 knee meeting inclusion criteria and 22 individuals with both knees meeting inclusion criteria. The cohort was 55% female with a mean age of 63.0 (SD 8.0) years and a mean BMI of 30.5 (SD 5.0) kg/m². We found that 24.6% of the cohort experienced significant knee pain based on the KOOS pain subscale and 13.4% reported experiencing a previous knee injury. The makeup of this analysis cohort mimics that of the overall OAI cohort. The baseline OAI cohort (n ​= ​4791 knees having undergone MRI) was 58% female with a mean age of 61.2 (SD 9.2) and a mean BMI of 28.6 (SD 4.8) kg/m² [25]. Distributions of JSW loss and MRI-based cartilage thickness loss measurements can be viewed in Supplemental Material.

Table 1.

Descriptive statistics of the analysis cohort (n ​= ​256).

	Entire Cohort n ​= ​256	Male n ​= ​114	Female n ​= ​142
Age mean ​± ​SD, years	63.0 ​± ​8.0	63.4 ​± ​8.4	62.6 ​± ​7.7
Missing n	0	0	0
BMI mean ​± ​SD, kg/m2	30.5 ​± ​5.0	29.7 ​± ​3.8	31.1 ​± ​5.6
Missing n	1	0	1
Race n (%)
White/Caucasian	212 (82.8)	96 (84.2)	116 (81.7)
African American	38 (14.8)	15 (13.2)	23 (16.2)
Asian	1 (0.4)	1 (0.9)	0 (0)
Other	5 (2.0)	2 (1.8)	3 (2.1)
Missing n	0	0	0
Ethnicity n (%)
Hispanic	3 (1.2)	1 (0.9)	2 (1.4)
Non-Hispanic	253 (98.8)	113 (99.1)	140 (98.6)
Missing n	0	0	0
KOOS Pain Score mean ​± ​SD	82.5 ​± ​19.0	85.3 ​± ​17.1	80.28 ​± ​20.2
Missing n	0	0	0
Categorized Pain n (%)
Painful	63 (24.6)	22 (19.3)	41 (28.9)
Non-painful	193 (75.4)	92 (80.7)	101 (71.1)
Previous Knee Injury n (%)
Yes	34 (13.4)	11 (9.7)	23 (16.3)
No	220 (86.6)	102 (90.3)	118 (83.7)
Missing n	2	1	1
JSW at T0a mean ​± ​SD, mm	6.0 ​± ​1.1	6.5 ​± ​1.1	5.6 ​± ​1.0
Missing n	2	2	0
JSW at T1b mean ​± ​SD, mm	4.8 ​± ​1.1	5.3 ​± ​1.2	4.4 ​± ​0.9
Missing n	0	0	0
Cartilage thickness (mm) in femoral condyle at T0a mean ​± ​SD, mm	1.9 ​± ​0.3	2.1 ​± ​0.3	1.8 ​± ​0.3
Missing n	0	0	0
Cartilage thickness (mm) in femoral condyle at T1b mean ​± ​SD, mm	1.7 ​± ​0.3	1.9 ​± ​0.3	1.6 ​± ​0.3
Missing n	0	0	0
Cartilage thickness (mm) in tibial condyle at T0a mean ​± ​SD, mm	1.7 ​± ​0.3	1.9 ​± ​0.3	1.6 ​± ​0.3
Missing n	0	0	0
Cartilage thickness (mm) in tibial condyle at T1b mean ​± ​SD, mm	1.6 ​± ​0.4	1.8 ​± ​0.4	1.5 ​± ​0.3
Missing n	0	0	0

^aIndicates the last follow-up timepoint where knees receive a JSN score of O.

^bIndicates the first follow-up timepoint where knees receive a JSN score of at least 1.

Pearson correlation coefficients for lnΔJSW and each MRI-based cartilage thickness measure are shown in Table 2. We found no or very weak correlations when assessing ΔMinFemTibCTh (R ​= ​0.00, p ​= ​0.98) and ΔMinFemCTh (R ​= ​0.03, p ​= ​0.61. There were weak, but stronger, correlations when examining ΔMeanFemCTh (R ​= ​0.26, p ​< ​0.01), ΔMeanTibCTh (R ​= ​0.23, p ​< ​0.01), and ΔMeanFemTibCTh (R ​= ​0.26, p ​< ​0.01). For each cartilage thickness measure, correlations were higher among females (n ​= ​142) compared to males (n ​= ​114). We did not observe a consistent pattern when stratifying results by painful (n ​= ​63) vs. non-painful knees (n ​= ​193). Correlations were higher among painful knees compared to non-painful knees when assessing ΔMeanFemCTh (R ​= ​0.33, p ​< ​0.01 vs. R ​= ​0.24, p ​< ​0.01), ΔMeanTibCTh (R ​= ​0.36, p ​< ​0.01 vs. R ​= ​0.20, p ​< ​0.01), and ΔMeanFemTibCTh (R ​= ​0.37, p ​< ​0.01) vs. R ​= ​0.23, p ​< ​0.01). Correlations were lower among painful knees compared to non-painful knees when assessing the remaining five MRI-based cartilage thickness measures.

Table 2.

Pearson correlation coefficients (R (p-value)) between natural log-transformed JSW loss (lnΔJSW) and each MRI-based cartilage thickness loss measure.

MRI measure	Entire Cohort (n ​= ​256)	Males (n ​= ​114)	Females (n ​= ​142)	Non-painful (n ​= ​193)	Painful (n ​= ​63)
ΔMinFemCTh	0.03 (0.61)	−0.13 (0.20)	0.12 (0.17)	0.09 (0.21)	−0.10 (0.45)
ΔMeanFemCTh	0.26 (<0.01)	0.22 (0.03)	0.30 (<0.01)	0.24 (<0.01)	0.33 (<0.01)
ΔMaxFemCTh	0.13 (0.04)	0.03 (0.75)	0.26 (<0.01)	0.16 (0.03)	0.03 (0.85)
ΔMeanTibCTh	0.23 (<0.01)	0.22 (0.02)	0.25 (<0.01)	0.20 (<0.01)	0.36 (<0.01)
ΔMaxTibCTh	0.17 (<0.01)	0.11 (0.27)	0.25 (<0.01)	0.20 (<0.01)	0.03 (0.83)
ΔMinFemTibCTh	0.00 (0.98)	−0.10 (0.32)	0.06 (0.50)	0.04 (0.56)	−0.09 (0.47)
ΔMeanFemTibCTh	0.26 (<0.01)	0.23 (0.02)	0.29 (<0.01)	0.23 (<0.01)	0.37 (<0.01)
ΔMaxFemTibCTh	0.17 (<0.01)	0.08 (0.40)	0.30 (<0.01)	0.20 (<0.01)	0.03 (0.82)

∗Note: ΔMinFemCTh is the measure of the minimum cartilage thickness loss within the femoral condyl; ΔMinFemTibCTh represents the combined femoral condyle and tibial plateau medial weight-bearing. ΔMeanFemCTh, ΔMeanTibCTh, and ΔMeanFemTibCTh represent the mean cartilage thickness loss within the femoral, tibial, and combined femoral and tibial weight-bearing region, respectively. ΔMaxFemCTh, ΔMaxTibCTh, and ΔMaxFemTibCTh is the measure of the maximum cartilage thickness loss within the femoral, tibial, and combined femoral and tibial weight-bearing region, respectively.

Linear mixed effect models show weak to no association (ranging from β ​= ​−0.02 to 0.25) between lnΔJSW and the majority of the MRI-based cartilage thickness loss measures among the entire analysis cohort and within sex and pain subgroups, and after adjusting for the previously mentioned covariates (Fig. 3, Fig. 4, Fig. 5 and Supplemental Figures 19-20 and 22-23). Estimated associations were slightly higher among females compared to males, particularly when the tibial cartilage was included. We found significantly higher associations between lnΔJSW and ΔMaxTibCTh, which represented maximum MRI-based cartilage thickness loss from the tibial plateau alone (ranging from β ​= ​3.41 to 4.02) (Supplemental Figure 21).

Fig. 3.

Results from linear mixed-effect models investigating the relationship between JSW loss (lnΔJSW) and mean MRI-based cartilage thickness loss from the medial weight-bearing region of the femoral condyle (ΔMeanFemCTh).

Fig. 4.

Results from linear mixed-effect models investigating the relationship between JSW loss (lnΔJSW) and mean MRI-based cartilage thickness loss from the medial weight-bearing region of the tibial plateau (ΔMeanTibCTh).

Fig. 5.

Results from linear mixed-effect models investigating the relationship between JSW loss (lnΔSW) and mean MRI-based cartilage thickness loss from the medial weight-bearing region of femoral condyle and tibial plateau (ΔMeanFemTibCTh).

4. Discussion

There has been increased debate recently around the potential for MRI-based measures of cartilage thickness to determine participant enrollment criteria and study endpoints in DMOAD clinical trials, which currently use radiographic-based measures for such standards [3,11,26,27]. While MRI offers direct visualization of cartilage thickness and health, radiography is a less expensive, less time-consuming, and more widely accessible imaging modality that is still deemed suitable in trial settings [28,29]. The fact remains that, despite decades of trial efforts, no DMOAD has been approved for use to slow or prevent further OA progression [27,30,31].

In this study, we sought to quantify the relationship between radiographic lnΔJSW, the standard measure recommended for DMOAD trial use, and MRI-based cartilage thickness loss, a potential measure to replace or be used in addition to JSW in trial settings. We also sought to stratify this relationship between the clinically meaningful subgroups of sex and pain status. We assessed these measures among knees likely in the early stages of OA—when they transitioned from JSN 0 (referred to as timepoint T₀) to JSN ≥1 (referred to as timepoint T₁). We found weak positive Pearson correlations between lnΔJSW and multiple versions of MRI-based cartilage thickness loss within the weight-bearing tibiofemoral joint. While still moderate in magnitude, the largest correlations were identified between lnΔJSW loss and mean MRI-based cartilage thickness loss. These results were comparable when using ΔMeanFemCTh and ΔMeanFemTibCTh, meaning that including additional data from the tibial plateau did not greatly affect this association. Similarly, a study by Lonza et al. which included a younger cohort of active individuals following anterior cruciate ligament reconstruction, found higher correlations between midpoint JSW and MRI-based cartilage thickness within the medial tibiofemoral compartment among females compared to males (R² ​= ​0.61 among females vs. R² ​= ​0.12 among males) [12].

On the other hand, linear mixed effect models showed generally weak associations between lnΔJSW and each MRI-derived cartilage thickness measure, indicating that JSW loss is not an ideal representation of cartilage loss in the weight-bearing region of the tibiofemoral joint. These results align with those of previous studies investigating the relationship between JSW and MRI-based cartilage thickness, both cross-sectionally and over time [14,23,32,33]. Estimates did not differ greatly between painful and non-painful knees. This finding makes sense given the general agreement that pain is not highly correlated with OA progression [16,34,35]. We found much higher associations between lnΔJSW and ΔMaxTibCTh loss for each cohort subgroup (ranging from β ​= ​3.41 to 4.02). Interpreting the β estimates as the mm change in lnΔJSW associated with 1 ​mm of change in ΔMaxTibCTh, results indicate that these two measures do not respond the same to joint changes that occur over time and within the medial weight-bearing joint region.

Associations were higher among females compared to males across all MRI-based cartilage thickness loss measures. These findings suggest that anatomic differences between male and female tibiofemoral joints may influence the relationship between ΔJSW and cartilage thickness loss. Such differences may include mean medial tibial slopes, which tend to be higher among females compared to males, or mean coronal tibial slopes, which are likely to be lower among females compared to males [36]. Another potential difference involves subtalar rotation—females are more likely to have higher subtalar joint protonation, and thus increased tibial rotation, compared to males [37]. Further work may be needed to fully understand which anatomic or biomechanical aspects account for the differences between biological sex we see in this study.

This study offers notable strengths. First, our restricted sample, which was defined by JSN transitions, allowed alignment of participants to a clinically meaningful baseline (T₀) and 12-month or 24-months of follow-up (T₁). Participants were enrolled in the OAI at various stages of knee OA, meaning that OAI baseline did not necessarily indicate that an individual was in early-stage disease—OAI participants may have been enrolled when they were already in late-stage disease, or they may have never developed knee OA throughout the study's follow-up period. The restricted sample allowed us to examine knees across the first eight years of follow-up that were likely in the early stages of OA, when progression may be slowed or prevented with proper treatment. We were also able to seamlessly compare 12-months or 24-months of progression across the included knees. Additionally, we chose to use JSN transitions in this way as medial JSN has been shown to be associated with greater future structural disease progression in the medial joint aspect [38]. Our analysis cohort (n ​= ​256 knees) was larger than the cohorts (ranging from n ​= ​23 to n ​= ​178 individuals with 1 included knee) in many of the previously published literature on this topic [12,14,33,39,40]. It should be noted, however, that Wirth et al. conducted a much larger study, with 967 knees from the OAI, to investigate the relationship between minimum JSW and MRI-based cartilage thickness change over 12–24 months [13].

We used MRI-based cartilage thickness data solely from the medial weight-bearing region of the tibiofemoral joint in this study, while other studies on this topic have used data from the entire medial and/or lateral joint aspects [12,14,33,41]. Cartilage thickness is known to vary across both the medial femoral condyle and medial tibial plateau [42]. We chose to focus on the medial weight-bearing region as it is known to experience greater cartilage volume loss compared to other regions and to limit the potential influence of thickness variation on results [23]. We found lower correlations between JSW loss and MRI-based cartilage thickness loss compared to the previously noted studies that focused on larger joint regions. While results from those studies indicate that JSW is a poor proxy measure for cartilage thickness, our findings in this study suggest that JSW is an even poorer surrogate measure when only focusing on a smaller, more localized joint region.

Historically, it has been difficult to obtain quantitative measures of cartilage thickness from MRI as this process involved manual segmentation techniques that could only be accomplished on a small subset of images [22]. Deep learning processes, such as the technique used to acquire this study's cartilage thickness data, can now effectively and efficiently analyze large image datasets to determine cartilage thickness within localized joint regions [22]. MRI, coupled with deep learning algorithms, may be a viable and effective tool for future DMOAD clinical trial settings. To the best of our knowledge, our study is the first to use MRI-based cartilage thickness data derived from deep learning processes in the context of comparing JSW to cartilage thickness in the tibiofemoral joint.

There are limitations associated with this study. Knee positioning during acquisition of MRI and radiographs could affect our results. Radiographs and MRI were collected using standardized protocols throughout the OA [17], but results may not be reproducible if using different imaging acquisition protocols (including a weight-bearing MRI protocol) or MRI data extraction methods. JSN scores are inherently limited as they are known to be affected by joint and meniscal positioning during imaging acquisition [10,43]. It is possible that knees included in this analysis may have transitioned back to a JSN score of 0 after meeting our specified inclusion criteria due to positioning errors. Finally, the OAI includes older adults and is comprised of predominantly White, non-Hispanic individuals. Our study may lack generalizability to younger, more active cohorts, as well as those belonging to other racial and ethnic groups.

5. Conclusions

Our study of the relationship between lnΔJSW and MRI-based cartilage thickness loss between early stages of JSN seems to quantitatively support the idea that JSW is not a sufficient proxy measure of cartilage thickness. There is potential for MRI-based measures of cartilage morphometry to be used instead of or in addition to standard radiographic measures of OA in DMOAD clinical trials [3,6,27,27]. As deep learning methods continue to increase the ease and efficiency of data extraction, MRI will likely become a more viable and accessible option to determine participant inclusion criteria and structural endpoints in trial settings.

Contributions

Substantial contributions to study conception and design: MCM, YMG, AEN. Substantial contributions to analysis and interpretation of data: MCM, TK, LA, MN, DN, JLL, SWM, JSM, YMG, AEN. Substantial drafting of the article: MCM, YMG, AEN. Critical revisions to the manuscript for intellectual content: MCM, TK, LA, MN, DN, JLL, SWM, JSM, YMG, AEN. Final approval of the manuscript: MCM, TK, LA, MN, DN, JLL, SWM, JSM, YMG, AEN.

Declaration of competing interest

Dr. Nelson reports receiving grants from NIH/NIAMS not related to this work, an honorarium from MedScape Education, membership on the Osteoarthritis Research Society International Board of Directors, and serving as an Associate Editor for Osteoarthritis and Cartilage.

Dr. Golightly receives grants from NIH/NIAMS not related to this work, is a member of the Association of Rheumatology Professionals Executive Committee, and serves as an Associate Editor for Osteoarthritis and Cartilage.

Role of the funding sources

MCM reports funding from the Thomas F. Ferdinand Summer Research Fellowship awarded through the Graduate School at the University of North Carolina and the Medical/Graduate Student Preceptorship awarded through the Rheumatology Research Foundation. The authors report funding from the Core Center for Clinical Research at the Thurston Arthritis Research Center. AEN reports support from the NIH/NIAMS (P30AR072580 and K24AR081368). JSM reports partial support by NSF Grant DMS-2113404. This work was supported in part by NIH grants 1R01AR072013 and 1R01AR082684. The work expresses the views of the authors, not of NIH.

Handling Editor: Professor H Madry

Appendix A. Supplementary data

The following is the Supplementary data to this article: