Abstract
Background
Bone marrow lesions (BMLs) are a known risk factor for incident knee osteoarthritis (OA), and deep learning (DL) methods can assist in automated segmentation and risk prediction.
Purpose
To develop and validate a DL model for quantifying tibiofemoral BML volume on MRI scans in knees without radiographic OA and to assess the association between longitudinal BML changes and incident knee OA.
Materials and Methods
This retrospective study included knee MRI scans from the Osteoarthritis Initiative prospective cohort (February 2004–October 2015). The DL model, developed between August and October 2023, segmented the tibiofemoral joint into 10 subregions and measured BML volume in each subregion. Baseline and 4-year follow-up MRI scans were analyzed. Knees without OA at baseline were categorized into three groups based on 4-year BML volume changes: BML-free, BML regression, and BML progression. The risk of developing radiographic and symptomatic OA over 9 years was compared among these groups.
Results
Included were 3869 non-OA knees in 2430 participants (mean age, 59.5 years ± 9.0 [SD]; female-to-male ratio, 1.3:1). At 4-year follow-up, 2216 knees remained BML-free, 1106 showed an increase in BML volume, and 547 showed a decrease in BML volume. BML progression was associated with a higher risk of developing radiographic knee OA compared with remaining BML-free (hazard ratio [HR] = 3.0; P < .001) or BML regression (HR = 2.0; P < .001). Knees with BML progression also had a higher risk of developing symptomatic OA compared with BML-free knees (HR = 1.3; P < .001). Larger volume changes in BML progression were associated with a higher risk of developing both radiographic OA (HR = 2.0; P < .001) and symptomatic OA (HR = 1.7; P < .001). In almost all subchondral plates, especially the medial femur and tibia, BML progression was associated with a higher risk of developing both radiographic and symptomatic OA compared with remaining BML-free.
Conclusion
Knees with BML progression, according to subregion and extent of volume changes, were associated with an increased risk of OA compared with BML-free knees and knees with BML regression, highlighting the potential utility of monitoring BML volume changes in evaluating interventions to prevent OA development.
ClinicalTrials.gov Identifier: NCT00080171
© RSNA, 2024
Supplemental material is available for this article.
See also the editorial by Said and Sakly in this issue.
Summary
Longitudinal increase in bone marrow lesion volume measured using an artificial intelligence algorithm on serial MRI scans was predictive of developing radiographic knee osteoarthritis.
Key Results
■ In a retrospective analysis of prospectively acquired knee MRI data (n = 9400 knees) from the Osteoarthritis Initiative, an automated deep learning model was developed and validated for bone marrow lesion (BML) segmentation and quantification.
■ Knees with BML progression over 4 years had a higher risk of developing radiographic (hazard ratio [HR] = 3.0; P < .001) and symptomatic (HR = 1.3; P < .001) osteoarthritis over 9 years, compared with BML-free knees.
Introduction
Bone marrow lesions (BMLs) are a key type of structural damage in knee osteoarthritis (OA), associated with altered biomechanics and reflecting bone remodeling (1,2). BMLs can appear early in OA development, even before radiographic or symptomatic evidence of OA (3–6). The role of BMLs in predicting symptomatic and structural OA progression, as well as knee replacement, has been investigated in knees with baseline OA (7). Not only is the presence of BMLs associated with increased risk for knee OA progression, but also changes in BML volume are associated with structural and symptomatic progression (5,8–10). Based on the clinical significance of BMLs, scoring systems like the MRI Osteoarthritis Knee Score (MOAKS) incorporate semiquantitative assessments of BML size across knee subregions (11–15) and are widely used in clinical trials and epidemiologic studies of knee OA (16–19).
While the link between BMLs and knee OA progression has been established, the role of BMLs in early OA stages is less clear. Roemer et al (20) demonstrated that the risk of OA development within 2 years was significantly elevated in knees with BMLs, particularly knees with medial BMLs. Additionally, in individuals without baseline knee OA, development of BMLs over a 2-year period was linked with a higher likelihood of developing knee pain (4). To determine the relationship between longitudinal changes in BML volume and downstream development of OA, quantitative BML measurements, robust analysis, larger sample sizes, extended follow-up periods, and meticulous adjustment for confounders are warranted.
Manual segmentation of knee MRI scans can be a labor-intensive task, and the inherent intra- and interreader variabilities challenge the credibility and consistency of BML segmentation. Deep learning (DL) algorithms present an opportunity for rapid and reliable segmentation of BMLs (21,22). Although there have been recent advancements in the development of fully automated DL models for BML segmentation, several aspects remain unexplored, including designing these models for subregional measurements, correlating model outputs with established semiquantitative methods like MOAKS, and using these models to examine the longitudinal relationship between BML volume and OA outcomes. Our purpose was to develop and validate a DL model for quantifying tibiofemoral BML volumes on MRI scans in knees without radiographic OA and to assess the association between longitudinal changes in BML volume and the development of OA.
Materials and Methods
Study Sample
Data were from the longitudinal multicenter Osteoarthritis Initiative (OAI) cohort study (February 2004 to October 2015; ClinicalTrials.gov identifier: NCT00080171). The OAI enrolled 4796 participants (45–79 years of age, from various ethnic groups) with or at risk of knee OA, from five clinical centers in the United States. The participants gave informed consent, and the study protocol was approved by the centers’ ethics committees (23). We used the OAI data set files listed in Table S1.
Exclusions in the OAI included individuals with MRI contraindications, inflammatory arthropathies, pregnancy, or bilateral knee replacement surgery. Further exclusions for this study were knees with baseline OA (Kellgren-Lawrence grade ≥ 2) and knees with missing MRI examination at baseline or 4-year follow-up (Fig 1).
Figure 1:
Flowchart of study selection criteria and categorization of included knees. BML = bone marrow lesion, KL = Kellgren-Lawrence, OA = osteoarthritis, OAI = Osteoarthritis Initiative.
Image Acquisition and Preprocessing
Knee MRI scans were obtained with 3-T Magnetom Trio systems (Siemens Healthineers) using a sagittal intermediate-weighted turbo spin-echo fat-suppressed sequence (repetition time msec/echo time msec, 3200/30; 0.357 × 0.357 × 3.0 mm). All MRI scans from baseline to year 4 in the OAI cohort were preprocessed (image conversion, reorientation, and resizing; Appendix S1).
DL Model Development
The DL model was developed between August and October 2023. Two readers manually segmented MRI scans to produce ground truth labels for DL model development and validation (Appendix S2). Model development (Fig 2) encompassed three distinct stages: (a) training the subregion model to segment the bones into 10 subregions akin to MOAKS (after the automated definition of the femur and tibia boundaries; Appendix S3), (b) training the BML model to segment BMLs within these tibiofemoral subregions, and (c) refining and quantifying the segmented BMLs (Appendix S4). We used two-dimensional U-Net architectures on the ResNet-50 backbone to develop the subregion and BML models with a fivefold cross-validation approach (details in Appendix S5). The Python codes used for DL model development are publicly accessible in the following GitHub repository: https://github.com/BMLSegmentation/Deep-learning-model-OAI.
Figure 2:
Diagram of the preprocessing and development of the two deep learning (DL) models and the segmentation of knee MRI scans. The process of measuring bone marrow lesion (BML) volume on knee MRI scans was divided into three phases. In phase I, a DL model was trained and validated to segment the tibiofemoral joint into 10 MRI Osteoarthritis Knee Score (MOAKS) subregions. In phase II, another DL model was trained and validated to segment subchondral BMLs. Finally, in phase III, the developed models were used to segment tibiofemoral joints on all knee MRI scans of the Osteoarthritis Initiative (OAI) cohort and to measure BML volume in each subregion. The external validity of the developed model was also assessed using the leave-one-center-out approach. BL = baseline, CV = cross-validation, IW = intermediate-weighted, MAT = medial anterior tibia, MCF = medial central femur, MCT = medial central tibia, MPF = medial posterior femur, MPT = medial posterior tibia, 2D = two-dimensional, Y-4 = year 4.
Internal and External Validation of the DL Models
Figure S1 shows an example of BML segmentation in each subregion. Three assessments were performed to test the validity of the model measurements. First, our models were internally validated using a fivefold cross-validation approach (24). Second, we externally validated our models by splitting the study sample by clinical center. We trained models on four centers and tested them on the remaining center, using a leave-one-center-out cross-validation approach with five iterations (25,26) (Appendix S6). The Dice similarity coefficient was used to measure the concordance between the manual and DL segmentations (27). Finally, using the Spearman correlation coefficient, we tested the correlation between the DL-derived quantitative BML volumes at baseline and the MOAKS-derived semiquantitative BML sizes (score range, 0–3). This analysis was conducted in 600 participants in the Osteoarthritis Biomarkers Project (FNIH Biomarkers Consortium), a nested OAI substudy with available MOAKS measurements.
Longitudinal Changes in BML Volume
Knees were categorized into three groups according to BML volume change from baseline to year 4 (exposure variable): no subchondral BMLs at baseline and year 4 (BML-free), decrease in BML volume or resolution of BMLs at follow-up (BML regression), and onset or increase in BMLs at follow-up (BML progression).
Knee OA Outcomes
Outcomes were the development of radiographic and symptomatic OA over 9 years in knees without OA at baseline (Kellgren-Lawrence grade < 2) (Fig 1). To prevent selection bias, we included all incident OA from baseline to year 9 in the analyses. We defined incident radiographic OA as development of Kellgren-Lawrence grade 2 or higher over an annual follow-up period. Similar to previous studies using OAI data (28,29), incident symptomatic OA was defined as the occurrence of frequent knee symptoms over the past year, defined as experiencing "pain, aching or stiffness in or around the knee on most days" for 1 month or more during the previous year.
Statistical Analysis
A radiology postdoctoral research fellow (K.M.) performed the statistical analyses using R version 4.0.3 packages (R Foundation for Statistical Computing). To examine between-group differences in baseline characteristics, we used the independent t test with the Levene test for equality of variances for continuous measures and the Fisher exact test for categorical measures. Three distinct analyses were performed to examine the association between the study exposure and outcomes while adjusting for age, sex, physical activity, and body mass index (calculated as weight in kilograms divided by height in meters squared). First, incident radiographic and symptomatic knee OA (binary time-to-event outcomes) were compared between the BML progression, BML regression, and BML-free groups using Cox proportional hazards analysis. The results are reported as hazard ratios (HRs) and 95% CIs. Second, using the same Cox models, we examined the association between the 4-year changes in BML volume within each subchondral bone plate and incident radiographic and symptomatic knee OA. Each subchondral bone plate of the tibiofemoral joint is composed of a combination of MOAKS subregions: medial femur (medial central femur and medial posterior femur), lateral femur (lateral central femur and lateral posterior femur), medial tibia (medial anterior tibia, medial central tibia, and medial posterior tibia), and lateral tibia (lateral anterior tibia, lateral central tibia, and lateral posterior tibia). Third, we determined whether the degree of BML volume increase was associated with the risk of OA development: Using maximally selected rank statistics (30), we identified a cutoff point for BML volume change to categorize the BML progression group into knees with large versus small BML changes. A Cox model was then used to compare incident OA between these two subgroups.
The false discovery rate for multiple comparisons was corrected using the Benjamini-Hochberg procedure. A two-tailed false discovery rate–corrected P < .05 was considered to represent a statistically significant difference.
Sensitivity Analysis
We performed sensitivity analyses to validate our results. First, to account for the potential prognostic factor bias arising from the time overlap between changes in BML volume (ie, exposure) and incident OA outcomes during the first 4 years of the study period (period 1), we analyzed the association between BML volume change during period 1 and the risk of developing radiographic and symptomatic OA in years 4–9 of the study period (period 2) using Cox models (Fig 3). Second, we stratified the data based on participants’ sex and repeated the analyses. Third, we categorized knees into the BML progression, BML regression, and BML-free groups based on 2-year (rather than 4-year) changes in BML volumes. Fourth, we addressed the issue that the DL algorithm used for calculating BML volume introduces a measurement error, such that subtle measurement errors, rather than actual changes, may have led several knees to be categorized as having BML progression or regression. To address this, we used the concept of the minimal detectable change (MDC), which is a statistical estimate of the smallest change detectable via a measure corresponding to a noticeable change in ability (31). It is calculated using the following formula:
Figure 3:
Diagram illustrating the timeline of study assessments for the main and sensitivity analyses, focusing on the association between changes in bone marrow lesion (BML) volume and the development of knee osteoarthritis (OA). Period 1 is baseline visit to 4-year follow-up visit. Period 2 is 4-year follow-up visit to 9-year follow-up visit. KL = Kellgren-Lawrence.
 |
where SEM is the standard error of the mean. We subsequently adjusted the knee categorization to incorporate a fourth group, unchanged BML volume, which included knees for which the volume change in BMLs was less than the MDC value. Fifth, the BML progression group comprised knees with an initially positive baseline BML value that subsequently increased and knees that presented no measurable BML value at baseline but developed BMLs by year 4. To address this variation, we stratified the knees into these two subgroups and conducted a comparative analysis of the time to development of OA.
Results
DL Model Performance
The segmentations from the DL models demonstrated a high level of agreement with the manual segmentations across the fivefold and leave-one-center-out cross-validation sets (Dice similarity coefficient range, 0.86–0.88; Table 1). Moreover, the BML volumes derived from the DL models strongly correlated with the MOAKS-derived semiquantitative scores across all subregions (Table 2). The medial posterior femur subregion had the lowest correlation coefficient (R = 0.80 [95% CI: 0.77, 0.82]), probably due to the cruciate ligament obscuring visibility into this area.
Table 1:
Assessment of the Internal and External Validity of the Developed Models
Table 2:
Validation of the U-Net Model Using Semiquantitative MOAKS Segmentation Results Derived from the Osteoarthritis Biomarkers Project (FNIH Biomarkers Consortium; n = 600 Knees)
Sample Characteristics
Of the 9592 knees from 4796 participants in the OAI cohort that were assessed for eligibility, 3869 knees in 2430 participants (mean age, 59.5 years ± 9.0 [SD]; female-to-male ratio, 1.3:1) were identified as not having knee OA at baseline and were included. Of the 3869 OA-free knees at baseline, 2216 remained BML-free over the subsequent 4-year follow-up period (BML-free group), 1106 had BML progression, and 547 had BML regression (Fig 1). Baseline characteristics of the knees are presented in Table 3.
Table 3:
Baseline Participant and Knee Characteristics of Knees Included in Study
Association of Longitudinal Changes in BML Volume with Development of Knee OA
As presented in Table 4, Cox analyses revealed that both BML progression and BML regression at 4-year follow-up were associated with a higher risk of radiographic knee OA compared with remaining BML-free (BML progression: HR = 3.0 [95% CI: 2.5, 3.6]; BML regression: HR = 1.6 [95% CI: 1.3, 2.1]). However, only BML progression was associated with a higher risk of symptomatic knee OA (HR = 1.3 [95% CI: 1.1, 1.4]). Moreover, knees with BML progression had a higher risk of radiographic OA (HR = 2.0 [95% CI: 1.6, 2.6]) but not symptomatic OA (HR = 1.1 [95% CI: 0.9, 1.3]) compared with knees with BML regression.
Table 4:
Association of Longitudinal Changes in BML Volume with Development of Radiographic and Symptomatic Knee OA
Prediction of Knee OA Based on the Degree of BML Volume Change
In knees with BML progression over 4 years, an increase of 1 SD in BML volume was associated with a higher risk of both radiographic OA (HR = 1.2 [95% CI: 1.2, 1.3]; P < .001) and symptomatic OA (HR = 1.1 [95% CI: 1.1, 1.2]; P < .001). Optimal cutoff points for 4-year change in BML volume were determined to be 85 mm3 for radiographic OA and 204 mm3 for symptomatic OA (Fig 4). BML progression exceeding these cutoff points was classified as a large BML change, while progression below these points was deemed a small BML change. In analysis adjusted for age, sex, body mass index, physical activity, and baseline BML volume, knees with large BML changes had a higher risk of both radiographic OA (HR = 2.0 [95% CI: 1.5, 2.5]) and symptomatic OA (HR = 1.7 [95% CI: 1.3, 2.3]) compared with those with small BML changes.
Figure 4:
Graphs showing the predictive value of large versus small changes in bone marrow lesion (BML) volume for (A) radiographic and (B) symptomatic knee osteoarthritis (OA). Using the maximally selected rank statistics method, we identified the cutoff points that maximized the survival difference between the two groups. Adjusted Cox proportional hazards models were then employed to compare the development of knee OA between these groups. BMI = body mass index (calculated as weight in kilograms divided by height in meters squared), HR = hazard ratio.
Prediction of Knee OA Based on BML Volume Change in Individual Subchondral Plates
We analyzed the association of 4-year BML volume change in the four subchondral plates (medial femur, lateral femur, medial tibia, and lateral tibia) with incident OA. Knees with increased BML volume in any plate had a higher risk of radiographic OA than BML-free knees (HR range, 1.8–2.5; Table 5). Furthermore, except for the lateral tibia, knees with BML progression in any plate had a higher risk of symptomatic OA (HR = 1.3 for medial femur, lateral femur, and medial tibia; Table 5). Among the four plates, the highest risk for knee OA was associated with BML progression in the medial subchondral plates, including the medial femur (radiographic OA: HR = 2.5; symptomatic OA: HR = 1.3) and the medial tibia (radiographic OA: HR = 2.5; symptomatic OA: HR = 1.3). Plate-specific analyses showed similar risk between knees with BML regression and BML-free knees (P > .05 for all comparisons; Table 5). Appendix S7 provides details on the association of BML volume change with incident knee OA in the 10 MOAKS subregions.
Table 5:
Association of Longitudinal Changes in BML Volume with Development of Radiographic and Symptomatic Knee OA According to Anatomic Subchondral Plate
Sensitivity Analysis
For the first sensitivity analysis, we examined the association of BML volume change in the first 4 years in the cohort (period 1) with incident OA during years 4–9 (period 2). Our results were consistent with the main analysis (Appendix S8). Second, we stratified the data by sex and found that BML volume change was associated with incident knee OA in both male and female participants (Appendix S9). Third, the results remained consistent after changing the period for observing changes in BML volume to 2 years (Appendix S10). Fourth, after changing the classification criteria to include a fourth group, unchanged BMLs, the association between BML volume change and incident OA remained consistent (Appendix S11). Fifth, when the BML progression group was subdivided according to whether BMLs were present or absent at baseline, an increase of 1 SD in BML volume was linked to a higher incidence of OA both in knees with initially positive baseline BML values and in those that presented no measurable BMLs at baseline, but the risk of OA was notably higher in knees that had initially positive baseline BML values (Appendix S12).
Discussion
In this study, we developed a deep learning algorithm to segment bone marrow lesions (BMLs). The algorithm performed well in validation tests and correlated with MRI Osteoarthritis Knee Score, a common method for grading BMLs. Implementing this algorithm, we showed that, in knees without baseline osteoarthritis (OA), BML volume increase was associated with higher risk of radiographic and symptomatic OA than remaining BML-free, and was also associated with a higher risk of radiographic OA than BML volume decrease. Further analysis showed that having BML progression in any plate, especially medial ones, increased the risk of radiographic OA compared with plates remaining BML-free. Larger BML volume increase was associated with an increased risk of both radiographic and symptomatic OA.
While the relationship between BML volume changes and knee OA progression is known, no study to our knowledge has explored this link in individuals without knee OA at baseline. Several studies align with our research, showing that incident BMLs increased knee pain risk (4) and that a 1-cm2 increase in BML size on MRI scans correlated with a rise in pain score (5). These studies, however, had important limitations, such as not measuring the quantitative three-dimensional volume of BMLs, having small sample sizes, having short follow-up periods, focusing on either symptomatic or radiographic knee OA, and having ambiguity about whether BML volume changes preceded the development of OA. Furthermore, the simultaneous evaluation of BMLs and radiographic OA exposed the readers to a lack of blinding regarding these features. Our study used a fully automated DL model to create a large MRI data set nested within the larger OAI, minimizing biases from nonblinded readers. Also, we evaluated the association between BML volume change and both symptomatic and radiographic knee OA, revealing that BML volume change was a stronger predictor of radiographic OA.
BML distribution across distinct knee subregions influences knee OA–related outcomes such as pain severity. Aso et al (32) demonstrated that changes in the size of medial tibiofemoral BMLs, as graded using the Whole-Organ Magnetic Resonance Imaging Score, were associated with changes in weight-bearing pain severity. In another study nested within the OAI, Roemer et al (20) reported that knees with medial BMLs had a higher risk of developing knee OA over a 2-year period, suggesting a subregion-based risk. Our study, the first to our knowledge to analyze subregion-based BML volume changes and incident knee OA, found associations in nearly all subregions, especially in the medial femur and tibia. In addition, we found that the degree of volume change also played a role in the risk of incident OA. Specifically, larger changes corresponded to a higher risk of developing knee OA. While existing evidence supports an association between large BML volume at baseline and knee OA outcomes (10), our study is the first to our knowledge to explore volume change extent as a predictor for incident knee OA.
This study highlights the importance of tracking BML volume changes as a predictor for knee OA and could be relevant in clinical trials. The findings open avenues for interventions to reduce BML volume in early disease stages, potentially reducing knee OA risk. Past efforts to decrease BMLs targeted subchondral bone turnover with interventions such as bisphosphonates (33) and strontium ranelate (34), or mitigated biomechanical stress through patellofemoral bracing (35) or high tibial osteotomy (36). Knee subchondroplasty (37) and minimally invasive approaches such as bone marrow concentrate and platelet product injections (38) have also been proposed, but no longitudinal randomized clinical trials are currently available to assess the potential of these treatment approaches.
This study had several limitations. First, our DL model was specific to the tibiofemoral joint, but BML changes in the patellofemoral joint have also been linked to knee OA symptoms (39). Future studies may include these subregions in DL models to potentially improve model performance. Second, we used the leave-one-center-out cross-validation method for external validation (25,26), but the same study protocol was used across all OAI recruitment centers, which means that our model may not be fully generalizable to other MRI protocols. Third, BMLs are linked not only to knee OA but also to spontaneous insufficiency fractures, osteonecrosis, and trauma (40). The existing DL algorithm, however, is unable to distinguish the underlying etiopathogenesis of a BML. Also, in our data set we cannot fully exclude misclassification and inclusion of BMLs not related to OA that are likely to be prevalent in the large OAI data set. As of now, no DL algorithm to our knowledge has been developed with the ability to make such a differentiation. Fourth, while BMLs are a key aspect of knee joint disease, they are only one feature. Future models could evaluate other MRI-related features, for a more comprehensive understanding of knee OA risk. Fifth, although the OAI also includes knee MRI scans in years 6 and 8 of follow-up, we evaluated the change in BML volume from baseline to only year 4 (or from baseline to year 2 in a sensitivity analysis). The aim here was to explore the predictive value of changes in BML volume for future, rather than concurrent, OA development. Future studies may consider knee MRI scans from these later time points to investigate the concurrent correlation between changes in BML volume and knee OA development. Finally, our model has performance constraints. We quantified BMLs using two-dimensional images and stacked them to compute the BML volume. However, the application of three-dimensional models could potentially enhance model performance in measuring BML volume. Future models should focus on enhancing the technical aspects of model development.
In conclusion, knees that demonstrate an increase in bone marrow lesion (BML) volume over time are associated with an increased risk of knee osteoarthritis (OA) compared with BML-free knees or knees that demonstrate a decrease in BML volume over time, highlighting the potential utility of monitoring BML volume changes in evaluating interventions to prevent the development of OA.
Acknowledgments
Acknowledgments
This project was conducted by the Osteoarthritis Initiative project investigators using publicly available Osteoarthritis Initiative data sets and was financially supported by the National Institutes of Health. The Osteoarthritis Initiative, a collaborative project between public and private sectors, includes five contracts N01-AR-2–2258, N01-AR-2–2259, N01-AR-2–2260, N01-AR-2–2261, and N01-AR-2–2262. Private funding partners were Merck Research Laboratories, Novartis, GlaxoSmithKline, and Pfizer. The results of this work do not necessarily reflect the opinions of the Osteoarthritis Initiative project investigators, the National Institutes of Health, or the private funding partners.
A.G. and S.D. are co-senior authors.
This research was supported by the National Institute on Aging (award number P01AG066603) and the National Institute of Arthritis and Musculoskeletal and Skin Diseases (award number R01AR079620-01).
Data sharing: Data generated or analyzed during the study are available from the corresponding author by request.
Disclosures of conflicts of interest: K.M. No relevant relationships. S. Mohammadi No relevant relationships. F.W.R. Grants or contracts to institution from Else Kröner-Fresenius-Stiftung, consulting fees from Grünenthal, member of the Program Committee (2021–2024) and Guidelines Committee (2024–present) of Osteoarthritis Research Society International, shareholder in Boston Imaging Core Lab, editor-in-chief of Osteoarthritis Imaging, and associate editor for Radiology. S. Momtazmanesh No relevant relationships. Q.H. No relevant relationships. H.A.I. No relevant relationships. D.J.H. No relevant relationships. A.G. Consultant to the editor for Radiology. S.D. No relevant relationships.
Abbreviations:
- BML
- bone marrow lesion
- DL
- deep learning
- HR
- hazard ratio
- MOAKS
- MRI Osteoarthritis Knee Score
- OA
- osteoarthritis
- OAI
- Osteoarthritis Initiative
References
- 1. Zanetti M , Bruder E , Romero J , Hodler J . Bone marrow edema pattern in osteoarthritic knees: correlation between MR imaging and histologic findings . Radiology 2000. ; 215 ( 3 ): 835 – 840 . [DOI] [PubMed] [Google Scholar]
- 2. Taljanovic MS , Graham AR , Benjamin JB , et al . Bone marrow edema pattern in advanced hip osteoarthritis: quantitative assessment with magnetic resonance imaging and correlation with clinical examination, radiographic findings, and histopathology . Skeletal Radiol 2008. ; 37 ( 5 ): 423 – 431 . [DOI] [PubMed] [Google Scholar]
- 3. Walsh DA , Sofat N , Guermazi A , Hunter DJ . Osteoarthritis bone marrow lesions . Osteoarthritis Cartilage 2023. ; 31 ( 1 ): 11 – 17 . [DOI] [PubMed] [Google Scholar]
- 4. Davies-Tuck ML , Wluka AE , Wang Y , English DR , Giles GG , Cicuttini F . The natural history of bone marrow lesions in community-based adults with no clinical knee osteoarthritis . Ann Rheum Dis 2009. ; 68 ( 6 ): 904 – 908 . [DOI] [PubMed] [Google Scholar]
- 5. Foong YC , Khan HI , Blizzard L , et al . The clinical significance, natural history and predictors of bone marrow lesion change over eight years . Arthritis Res Ther 2014. ; 16 ( 4 ): R149 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6. Davies-Tuck ML , Wluka AE , Forbes A , et al . Development of bone marrow lesions is associated with adverse effects on knee cartilage while resolution is associated with improvement—a potential target for prevention of knee osteoarthritis: a longitudinal study . Arthritis Res Ther 2010. ; 12 ( 1 ): R10 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7. Roemer FW , Kwoh CK , Hannon MJ , et al . Can structural joint damage measured with MR imaging be used to predict knee replacement in the following year? Radiology 2015. ; 274 ( 3 ): 810 – 820 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8. Roemer FW , Guermazi A , Javaid MK , et al . Change in MRI-detected subchondral bone marrow lesions is associated with cartilage loss: the MOST Study. A longitudinal multicentre study of knee osteoarthritis . Ann Rheum Dis 2009. ; 68 ( 9 ): 1461 – 1465 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9. Zhang Y , Nevitt M , Niu J , et al . Fluctuation of knee pain and changes in bone marrow lesions, effusions, and synovitis on magnetic resonance imaging . Arthritis Rheum 2011. ; 63 ( 3 ): 691 – 699 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10. Driban JB , Price L , Lo GH , et al . Evaluation of bone marrow lesion volume as a knee osteoarthritis biomarker—longitudinal relationships with pain and structural changes: data from the Osteoarthritis Initiative . Arthritis Res Ther 2013. ; 15 ( 5 ): R112 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11. Jaremko JL , Jeffery D , Buller M , et al . Preliminary validation of the Knee Inflammation MRI Scoring System (KIMRISS) for grading bone marrow lesions in osteoarthritis of the knee: data from the Osteoarthritis Initiative . RMD Open 2017. ; 3 ( 1 ): e000355 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12. Roemer FW , Collins J , Kwoh CK , et al . MRI-based screening for structural definition of eligibility in clinical DMOAD trials: Rapid OsteoArthritis MRI Eligibility Score (ROAMES) . Osteoarthritis Cartilage 2020. ; 28 ( 1 ): 71 – 81 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13. Hunter DJ , Guermazi A , Lo GH , et al . Evolution of semi-quantitative whole joint assessment of knee OA: MOAKS (MRI Osteoarthritis Knee Score) . Osteoarthritis Cartilage 2011. ; 19 ( 8 ): 990 – 1002 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14. Hunter DJ , Lo GH , Gale D , Grainger AJ , Guermazi A , Conaghan PG . The reliability of a new scoring system for knee osteoarthritis MRI and the validity of bone marrow lesion assessment: BLOKS (Boston Leeds Osteoarthritis Knee Score) . Ann Rheum Dis 2008. ; 67 ( 2 ): 206 – 211 . [DOI] [PubMed] [Google Scholar]
- 15. Lynch JA , Roemer FW , Nevitt MC , et al . Comparison of BLOKS and WORMS scoring systems part I. Cross sectional comparison of methods to assess cartilage morphology, meniscal damage and bone marrow lesions on knee MRI: data from the Osteoarthritis Initiative . Osteoarthritis Cartilage 2010. ; 18 ( 11 ): 1393 – 1401 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16. Roemer FW , Guermazi A , Hannon MJ , et al . Presence of magnetic resonance imaging-defined inflammation particularly in overweight and obese women increases risk of radiographic knee osteoarthritis: the POMA study . Arthritis Care Res (Hoboken) 2022. ; 74 ( 8 ): 1391 – 1398 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17. Katz JN , Chaisson CE , Cole B , et al . The MeTeOR trial (Meniscal Tear in Osteoarthritis Research): rationale and design features . Contemp Clin Trials 2012. ; 33 ( 6 ): 1189 – 1196 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18. Roemer FW , Hunter DJ , Crema MD , Kwoh CK , Ochoa-Albiztegui E , Guermazi A . An illustrative overview of semi-quantitative MRI scoring of knee osteoarthritis: lessons learned from longitudinal observational studies . Osteoarthritis Cartilage 2016. ; 24 ( 2 ): 274 – 289 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19. Sandhar S , Smith TO , Toor K , Howe F , Sofat N . Risk factors for pain and functional impairment in people with knee and hip osteoarthritis: a systematic review and meta-analysis . BMJ Open 2020. ; 10 ( 8 ): e038720 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20. Roemer FW , Kwoh CK , Hannon MJ , et al . What comes first? Multitissue involvement leading to radiographic osteoarthritis: magnetic resonance imaging-based trajectory analysis over four years in the osteoarthritis initiative . Arthritis Rheumatol 2015. ; 67 ( 8 ): 2085 – 2096 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21. Preiswerk F , Sury MS , Wortman JR , Neumann G , Wells W , Duryea J . Fast quantitative bone marrow lesion measurement on knee MRI for the assessment of osteoarthritis . Osteoarthr Cartil Open 2022. ; 4 ( 1 ): 100234 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22. Dijkstra AJ , Anbeek P , Yang KG , et al . Validation of a novel semiautomated segmentation method for MRI detection of cartilage-related bone marrow lesions . Cartilage 2010. ; 1 ( 4 ): 328 – 334 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23. Nevitt M , Felson D , Lester G . The Osteoarthritis Initiative: protocol for the cohort study . Osteoarthritis Initiative . https://nda.nih.gov/static/docs/StudyDesignProtocolAndAppendices.pdf. Published June 21, 2006. Accessed December 19, 2023 .
- 24. Bengio Y , Grandvalet Y . No unbiased estimator of the variance of K-fold cross-validation . In: Thrun S , Saul L , Schölkopf , eds. Advances in neural information processing systems 16 (NIPS 2003) . https://proceedings.neurips.cc/paper/2003/hash/e82c4b19b8151ddc25d4d93baf7b908f-Abstract.html. Published 2003. Accessed December 19, 2023 . [Google Scholar]
- 25. Steyerberg EW , Harrell FE Jr . Prediction models need appropriate internal, internal-external, and external validation . J Clin Epidemiol 2016. ; 69 : 245 – 247 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26. König IR , Malley JD , Weimar C , Diener HC , Ziegler A ; German Stroke Study Collaboration . Practical experiences on the necessity of external validation . Stat Med 2007. ; 26 ( 30 ): 5499 – 5511 . [DOI] [PubMed] [Google Scholar]
- 27. Zou KH , Warfield SK , Bharatha A , et al . Statistical validation of image segmentation quality based on a spatial overlap index . Acad Radiol 2004. ; 11 ( 2 ): 178 – 189 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28. Riddle DL , Kong X , Jiranek WA . Two-year incidence and predictors of future knee arthroplasty in persons with symptomatic knee osteoarthritis: preliminary analysis of longitudinal data from the Osteoarthritis Initiative . Knee 2009. ; 16 ( 6 ): 494 – 500 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29. Laslett LL , Kingsbury SR , Hensor EMA , Bowes MA , Conaghan PG . Effect of bisphosphonate use in patients with symptomatic and radiographic knee osteoarthritis: data from the Osteoarthritis Initiative . Ann Rheum Dis 2014. ; 73 ( 5 ): 824 – 830 . [DOI] [PubMed] [Google Scholar]
- 30. Lausen B , Schumacher M . Maximally selected rank statistics . Biometrics 1992. ; 48 ( 1 ): 73 – 85 . [Google Scholar]
- 31. Beckerman H , Roebroeck ME , Lankhorst GJ , Becher JG , Bezemer PD , Verbeek AL . Smallest real difference, a link between reproducibility and responsiveness . Qual Life Res 2001. ; 10 ( 7 ): 571 – 578 . [DOI] [PubMed] [Google Scholar]
- 32. Aso K , Shahtaheri SM , McWilliams DF , Walsh DA . Association of subchondral bone marrow lesion localization with weight-bearing pain in people with knee osteoarthritis: data from the Osteoarthritis Initiative . Arthritis Res Ther 2021. ; 23 ( 1 ): 35 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33. Seefried L , Genest F , Baumann J , Heidemeier A , Meffert R , Jakob F . Efficacy of zoledronic acid in the treatment of nonmalignant painful bone marrow lesions: a triple‐blind, randomized, placebo‐controlled phase III clinical trial (ZoMARS) . J Bone Miner Res 2022. ; 37 ( 3 ): 420 – 427 . [DOI] [PubMed] [Google Scholar]
- 34. Pelletier J-P , Roubille C , Raynauld J-P , et al . Disease-modifying effect of strontium ranelate in a subset of patients from the phase III knee osteoarthritis study SEKOIA using quantitative MRI: reduction in bone marrow lesions protects against cartilage loss . Ann Rheum Dis 2015. ; 74 ( 2 ): 422 – 429 . [DOI] [PubMed] [Google Scholar]
- 35. Callaghan MJ , Parkes MJ , Hutchinson CE , et al . A randomised trial of a brace for patellofemoral osteoarthritis targeting knee pain and bone marrow lesions . Ann Rheum Dis 2015. ; 74 ( 6 ): 1164 – 1170 . [DOI] [PMC free article] [PubMed] [Google Scholar]
- 36. Choi HG , Kim JS , Yoo HJ , Jung YS , Lee YS . The fate of bone marrow lesions after open wedge high tibial osteotomy: a comparison between knees with primary osteoarthritis and subchondral insufficiency fractures . Am J Sports Med 2021. ; 49 ( 6 ): 1551 – 1560 . [DOI] [PubMed] [Google Scholar]
- 37. Chua K , Kang JYB , Ng FDJ , et al . Subchondroplasty for bone marrow lesions in the arthritic knee results in pain relief and improvement in function . J Knee Surg 2021. ; 34 ( 6 ): 665 – 671 . [DOI] [PubMed] [Google Scholar]
- 38. Centeno C , Cartier C , Stemper I , et al . The treatment of bone marrow lesions associated with advanced knee osteoarthritis: comparing intraosseous and intraarticular injections with bone marrow concentrate and platelet products . Pain Physician 2021. ; 24 ( 3 ): E279 – E288 . [PubMed] [Google Scholar]
- 39. Zhu Z , Ding C , Jin X , et al . Patellofemoral bone marrow lesions: natural history and associations with pain and structure . Arthritis Care Res (Hoboken) 2016. ; 68 ( 11 ): 1647 – 1654 . [DOI] [PubMed] [Google Scholar]
- 40. Gobbi A , Dallo I , Frank RM , Bradsell H , Saenz I , Murrel W . A review of bone marrow lesions in the arthritic knee and description of a technique for treatment . J Cartil Jt Preserv . 2021. ; 1 ( 3 ): 100021 . [Google Scholar]