HAL Id: hal-02047854
https://hal.archives-ouvertes.fr/hal-02047854
Submitted on 25 Feb 2019
HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
of radiographic knee osteoarthritis: data from the Osteoarthritis Initiative
Thomas Janvier, Rachid Jennane, Hechmi Toumi, Eric Lespessailles
To cite this version:
Thomas Janvier, Rachid Jennane, Hechmi Toumi, Eric Lespessailles. Subchondral tibial bone texture predicts the incidence of radiographic knee osteoarthritis: data from the Osteoarthritis Initiative.
Osteoarthritis and Cartilage, Elsevier, 2017, 25 (12), pp.2047-2054. �10.1016/j.joca.2017.09.004�. �hal- 02047854�
Subchondral Tibial Bone Texture Predicts the Incidence of Radiographic Knee Osteoarthritis: Data from the Osteoarthritis Initiative
T. Janvier, (PhD), R. Jennane, (PhD), H. Toumi, (PhD), E. Lespessailles, (MD, PhD)
PII: S1063-4584(17)31200-1 DOI: 10.1016/j.joca.2017.09.004 Reference: YJOCA 4087
To appear in: Osteoarthritis and Cartilage Received Date: 15 February 2017
Revised Date: 1 September 2017 Accepted Date: 8 September 2017
Please cite this article as: Janvier T, Jennane R, Toumi H, Lespessailles E, Subchondral Tibial Bone Texture Predicts the Incidence of Radiographic Knee Osteoarthritis: Data from the Osteoarthritis Initiative, Osteoarthritis and Cartilage (2017), doi: 10.1016/j.joca.2017.09.004.
This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
M AN US CR IP T
AC CE PT ED
Subchondral Tibial Bone Texture
1
Predicts the Incidence of Radiographic
2
Knee Osteoarthritis: Data from the
3
Osteoarthritis Initiative
4
T. Janvier (PhD) a, R. Jennane (PhD) a, H.Toumi (PhD) a,b, E. Lespessailles (MD, PhD) a,b*
5
a Univ. Orléans, I3MTO Laboratory, EA 4708, 45067 Orléans, France 6
b CHR Orléans, Rheumatology Department, 45032 Orléans, France 7
* Corresponding author:
8
Eric Lespessailles (MD, PhD) 9
Regional Hospital of Orleans, 14 avenue de l’hopital, 45067 Orleans Cedex 2 10
University of Orleans, I3MTO Laboratory, EA 4708, 45067 Orleans, France 11
Office: +33(0)238613151 12
E-mail : eric.lespessailles@chr-orleans.fr 13
Abstract
14
Objectives: To evaluate whether trabecular bone texture (TBT) parameters measured on 15
computed radiographs could predict the onset of radiographic knee osteoarthritis (OA).
16
Materials and Methods: Subjects from the Osteoarthritis Initiative with no sign of 17
radiographic OA at baseline were included. Cases that developed either a global radiographic 18
OA defined by the Kellgren-Lawrence (KL) scale, a joint space narrowing (JSN) or tibial 19
osteophytes (TOS) were compared with the controls with no changes after 48 months of 20
follow-up. Baseline bilateral fixed flexion computed radiographs were analyzed using a 21
M AN US CR IP T
AC CE PT ED
fractal method to characterize the local variations. The prediction was explored using logistic 22
regression models evaluated by the area under the receiver operating characteristic curves 23
(AUC).
24
Results: From the 344 knees, 79 (23%) developed radiographic OA after 48 months, 44 25
(13%) developed progressive JSN and 59 (17%) developed osteophytes. Neither age, gender 26
and BMI, nor their combination predicted poorer KL (AUC 0.57), JSN or TOS (AUC 0.59) 27
scores. The inclusion of the TBT parameters in the models improved the global prediction 28
results for KL (AUC 0.69), JSN (AUC 0.73) and TOS (AUC 0.71) scores.
29
Conclusions: Several differences were found between the models predictive of three 30
different outcomes (KL, JSN and TOS), indicating different underlying mechanisms. These 31
results suggest that TBT parameters assessed when radiographic signs are not yet apparent on 32
radiographs may be useful in predicting the onset of radiological tibiofemoral OA as well as 33
identifying at-risk patients for future clinical trials.
34
35
Keywords: fractal analysis, trabecular bone texture, knee osteoarthritis, subchondral bone, 36
radiographs 37
M AN US CR IP T
AC CE PT ED
Introduction
38
Although osteoarthritis (OA) is the most common disorder of the musculoskeletal system, 39
there is still no cure and treatment options are limited, leading to an increasing number of 40
total joint replacements 1,2 worldwide. The underlying mechanisms through which this 41
debilitating disease 1 occurs and progresses have not been fully elucidated yet 3,4. Although 42
the progression of the disease has been widely studied using several modalities 5–9, the 43
genesis of the pathology has been much less investigated 10,11. 44
Considering the growing body of evidence that showed the role of the subchondral bone 45
3,7,12,13
and the concept of crosstalk between articular cartilage and subchondral bone tissue 13, 46
we investigated the ability to predict the onset of knee OA based on observation of the 47
subchondral bone. Trabecular bone texture (TBT) has been used for a long time to 48
characterize altered trabecular bone 14 and is recognized to be a promising imaging biomarker 49
15. Moreover, TBT parameters (TBTp) including fractal descriptors have been consistently 50
reported to be able to predict the progression of radiographic knee OA 5,15–17 but also to be 51
associated with the risk of total knee joint replacement 18. Although conventional radiography 52
has been considered unable to identify early relevant cartilage and bone changes associated 53
with the onset of OA, TBT analysis could provide a second life to digital radiography and 54
should provide new interesting information on the pathogenesis and occurrence of knee OA 55
development 19. 56
M AN US CR IP T
AC CE PT ED
Therefore, in the present longitudinal study conducted over 48 months of participants in the 57
Osteoarthritis Initiative (OAI) we analyzed the predictive ability of TBTp measured on 58
computed radiographs to predict the incidence of radiographic knee OA. There are very few 59
studies 11,19,20 that have investigated the genesis of radiographic knee OA through the TBT, 60
yet, they showed that texture parameters combined with machine learning could predict knee 61
OA initiation. Our contribution resides in the confirmation of these results on a large and well 62
phenotyped cohort (OAI), using a reduced number of features reflecting the bone topology 63
and considering a short-term outcome. A statistical approach robust to overfitting using 64
cross-validation and a parsimonious reduction of the models was applied. Contrary to 19, 65
subjects with doubtful OA (KL=1) were excluded in the present work in order to focus on 66
healthy at risk patients. In 11,20, they considered a long range outcome (20 years follow-up) 67
while this study focuses on a short range outcome with a 4-year prediction. This work intends 68
to identify patients with an elevated short-term risk to develop radiographic knee OA.
69
Material & Methods
70
Patients 71
The data used for the present study were obtained from the OAI database 21, which is 72
available for public access at http://www.oai.ucsf.edu/. This database contains bilateral X-ray 73
M AN US CR IP T
AC CE PT ED
recruited in 5 different centers and had annual bilateral fixed flexion knee radiographs using 75
different equipment and modalities among the centers but all following the same acquisition 76
protocol using the fixed flexion positioner SynaFlexer 22. For this study, the radiographic 77
scoring used was provided by the Boston University X-ray reading center 23. This assessment 78
dataset includes both the Kellgren-Lawrence (KL) 24 and the Osteoarthritis Research Society 79
International (OARSI) 25 semi-quantitative scores for radiographs from baseline to month 48, 80
limiting the analyses to the 3421 graded knees in 1789 patients. As the present work focused 81
on the initiation of knee OA, we excluded the prevalent doubtful OA knees (e.g KL = 1 at 82
baseline), resulting in a subset of 514 knees from 471 patients. No standard protocols were 83
provided for digitization of the film radiographs, resulting in considerable variability among 84
the digitized images (variations in resolution, dynamic range, field of view, alignment, etc.) 85
and they were therefore excluded.
86
The final subset included 344 knees in 319 patients. Only computed radiographs (CR) were 87
analyzed. Digital radiographs (DR) were not considered in this study as none was available in 88
the OAI at baseline. The complete pruning process is detailed in Figure 1 and the resulting 89
population characteristics are detailed in Table 1.
90
OA grouping 91
Herein, the diagnosis of radiographic knee OA was defined as a KL score ≥ 1 according to 92
the OAI definition 21. The onset of OA was investigated in patients with initial KL grade 0 93
M AN US CR IP T
AC CE PT ED
(i.e. both joint space narrowing (JSN) and tibial osteophyte (TOS) grade 0) using three 94
definitions based on these scores and described in Table 1. Knees with a KL going from 0 at 95
baseline to 1 at 48 months (16 samples) were included as controls in order not to bias the 96
results compared to a real case scenario. However, the same computations were achieved 97
after removing these 16 samples, confirming the results presented here with non-significantly 98
higher predictive values (data not shown).
99
Processing of radiographs 100
Radiographs were provided by the OAI as anonymous 16-bit DICOM images from which 101
information such as the modality and the resolution were extracted. Each radiograph (right 102
and left knees) was evaluated independently. Based on the results obtained in a previous 103
study on the impact of ROI on OA characterization 26 we chose to use the whole tibial bone 104
area. The ROI extraction was performed by the principal investigator of the study, T. Janvier, 105
as detailed in our previous study 27. The resulting patchwork of 16 ROIs is presented in 106
Figure 2.
107
Each ROI side length was proportional to the tibia width to avoid misplacement due to 108
anatomical variations in the knee widths. In our subset, radiographs presented different pixel 109
spacing 150; 158; 169; 170; 200 µm without significant differences between groups, 110
giving ROI side lengths ranging from 9 to 11 mm.
111
M AN US CR IP T
AC CE PT ED
Trabecular bone texture analysis 112
Fractal analysis has been regularly used in the past for TBT characterization 15,18,28–30
. It has 113
been shown to be linked to the 3D microarchitecture of the bone 31–34 and to be robust to 114
different common artifacts of X-ray imaging such as radiographic magnification, projection 115
35,36
and inhomogeneous contrast 31,35. For these reasons, images were processed directly 116
from the DICOM files without any preprocessing.
117
In the present study, the fractal parameter (H) was computed using the variogram 37, which 118
gives a measure of how much two pixels from a radiograph will vary in intensity depending 119
on the distance between them. The H parameter is computed as the slope of the variogram in 120
a log-log scale and quantifies the intuitive notion of the roughness of a texture (i.e.
121
complexity of the texture evolution with the scale of observation).
122
The piecewise fractal nature of the TBT reported in 38,39 and confirmed by the variogram 123
shapes in Figure 3 was also taken into account. The cut-off scale was observed around 500 124
µm on the empirical variograms and two fractal parameters were extracted: and 125
corresponding to the texture complexity computed for the two micro (µ-scale) and the milli 126
(m-scale) scales of observation respectively under 400 µm and above 600 µm. These two 127
parameters ; were computed along both the vertical (90°) and horizontal (0°) axes 128
corresponding to the main loading compression and tension directions, respectively. Finally, 129
M AN US CR IP T
AC CE PT ED
TBTp represent a vector of 4 descriptors |°; |°; |°; |° computed on 16 ROIs 130
(Figure 2), resulting in 64 scalar parameters.
131
Statistical analysis 132
We compared the clinical covariates (age, gender, BMI) in patients with and without incident 133
radiographic knee OA using the Mann-Whitney and Chi-square tests for numerical and 134
categorical factors, respectively.
135
The predictive ability of TBTp was evaluated using a logistic regression through different 136
models including combinations of the clinical covariates and the TBTp. To avoid the bias 137
induced by a large number of parameters, these models were reduced using an automatic 138
iterative stepwise optimization algorithm based on the Akaike Information Criterion (age, 139
gender and BMI were not forced to be included in the final models). To avoid overfitting 140
effects, the models used were evaluated using a 10-fold cross-validation repeated 300 times.
141
The performance of each model was evaluated through the receiver-operating characteristic 142
(ROC) curves using the area under the curve (AUC) as a global performance criterion. The 143
diagnostic odd ratio (DOR) was also computed 40 to investigate the relevance of the different 144
models. DOR provides the ratio between the odds of a case being well classified against the 145
odds of a control being wrongly classified as a case. It focuses on patients predicted as 146
initiators and indicates if they were indeed true positives (DOR > 1) or false positives (DOR 147
M AN US CR IP T
AC CE PT ED
< 1). This indicator is used to characterize the relative risk of actual radiographic knee OA 148
incidence in a subpopulation predicted as initiators.
149
All statistical analyses were performed using R 41 (version 3.1.2) and the packages MASS 42 150
(version 7.3), Caret 43 (version 6.0) and pROC 44 (version 1.8).
151
Results
152
There were no differences at baseline between controls and patients with incident 153
radiographic knee OA in their clinical covariates (age, gender and BMI, in Table 1) for any 154
of the three definitions (∆JSN, ∆TOS, ∆KL) tested.
155
Table 2 presents the details of the significant associations of the parsimonious models’
156
features with knee OA initiation for the three definitions. The clustering effects due to the 157
inclusion of both knees were tested using the generalized estimating equations (GEE) 45. The 158
clusters were defined using the patient ID from the OAI dataset. These results show that 159
considering both knees from each subject helps to better explain the coefficients of the 160
models (JSN, TOS, KL), as parameters considered not significant with a naïve interpretation 161
become significant.
162
Figure 4 presents the ROC curves averaged for the 300 cross-validated runs and Table 2 163
details the AUC and DOR with 95% CI of the different predictive models. For each 164
definition, results of the models including the clinical covariates with and without the TBTp 165
M AN US CR IP T
AC CE PT ED
are shown. In Table 2, the results for partial models including the four specific features 166
|°; |°; |°; |° individually are detailed.
167
Joint Space Narrowing 168
The clinical covariates alone were not able to predict the increased JSN. Individual TBT 169
parameters computed in the vertical direction increased the results (AUC > 0.6), yet the 170
prediction was not suitable to discriminate patients with a higher risk of knee OA onset (DOR 171
< 1). TBTp computed in the horizontal direction enabled the prediction (AUC 0.7), however, 172
the prediction was found relevant (DOR > 1) only for the larger scales (m-scale). The 173
combination of all the texture information provided the best prediction (AUC 0.73) with a 174
higher risk for the patients highlighted as at risk (DOR 4.51).
175
Tibial Osteophytes 176
The clinical covariates alone were not able to predict the occurrence of TOS. Individual TBT 177
parameters computed in either vertical or horizontal direction increased the results (AUC >
178
0.6) with relevant potential selection of the patients at risk (DOR > 1) except for the larger 179
scale (m-scale) in vertical orientation. The combination of all the texture information 180
provided the best prediction (AUC 0.71) with a higher risk for the patients highlighted as at 181
risk (DOR 7.10).
182
M AN US CR IP T
AC CE PT ED
Radiographic knee OA 183
The clinical covariates alone were not able to predict the occurrence of radiographic knee OA 184
(KL ≥ 2). Individual TBT parameters computed in either vertical or horizontal direction 185
increased the results (AUC > 0.6), yet the prediction was relevant (DOR > 1) only for the 186
horizontal analysis. The combination of all the texture information provided the best 187
prediction (AUC 0.69) with a higher risk for the patients highlighted as at risk (DOR 2.33).
188
Discussion
189
In this study, we examined the ability of baseline TBTp measured on computed radiographs 190
in patients from the large cohort of the OAI to predict the incidence of radiographic knee OA 191
over a period of four years. Our findings confirmed that TBTp on baseline radiographs 192
predicted later onset of radiographic knee OA.
193
Recently, Podsiadlo et al. 19 reported that directional fractal signature analysis was associated 194
with the incidence of radiographic knee OA. The work by Podsiadlo and ours demonstrated 195
the potential interest of selecting a subgroup of patients having baseline characteristic 196
features of TBTp on plain radiographs prone to the development of radiographic knee OA. In 197
their work, 2052 knees were investigated with a KL grade ≤ 1 from 2 centers analyzed 198
individually, focusing on a single pair of ROIs using several fractal descriptors. However, 199
they included OA doubtful subjects (KL grade 1) in their baseline cohort. In the present 200
M AN US CR IP T
AC CE PT ED
work, only the non-doubtful healthy subjects at baseline were included, as advocated by the 201
OAI 21. 202
Our approach included the whole proximal tibial trabecular bone, mapping the area with 16 203
different ROIs to cover the global aspect of knee OA as defined by Loeser 3. We have already 204
demonstrated that relevant information to detect the risk of OA progression is provided by 205
both lateral and distal subchondral bone regions at the tibia and not only by the medial one 206
27,46
. 207
There were no significant differences between the common covariates (age, gender, BMI) at 208
baseline between patients with incident knee OA and controls at 4 years. BMI has been 209
shown to be positively associated with the onset of knee OA 47–49. However, BMI seems to be 210
more associated with knee pain 50 than with joint failure 51. Conversely, we showed that the 211
use of the TBTp could enable the prediction of incident radiographic knee OA.
212
Details of the TBT analysis 213
The differentiation between three incidence definitions (∆JSN, ∆TOS and ∆KL) and the dual 214
scale analysis showed that different results are obtained depending on the scenario.
215
First, it seems that the early changes induced by knee OA are observable along the horizontal 216
direction as reported by Buckland-Wright. These results are in line with previous data shown 217
by Podsiadlo et al. 19 where they found better results with a larger scale in computed 218
M AN US CR IP T
AC CE PT ED
the vertical direction for the progression of pre-existing knee OA. This implies that TBT 220
analysis is relevant in the horizontal direction for early diagnosis and then in the vertical one 221
for late follow-up. Extending this observation to the actual topology of the microarchitecture 222
of the trabecular bone implies that the OA process involves changes in the vertical trabeculae 223
(characterized by horizontal analysis) in the early stages of knee OA incidence, while in later 224
stages of knee OA progression these changes may occur mostly on the horizontal trabeculae 225
(characterized with a vertical analysis).
226
Second, the bifractal aspect of the TBT was arbitrarily chosen and delimited in this study 227
according to empirical observation of the variograms. However, the cut-off at 500 µm 228
corresponds to the breaking point in scale range between the trabecular thickness (100-300 229
µm) and the trabecular separation (200-2000 µm) 52. This observation has been reported in 230
several studies focusing on the TBT 15,17,38 and might confirm the need to differentiate the 231
process occurring at different scales in bone remodeling. Moreover, the decomposition into 232
partial models for the different definitions showed, as for the orientations, several differences.
233
The TOS seems to have more links with the global TBT features as three of the four 234
parameters provided relevant results, whereas the higher impact of large scale changes for 235
JSN prediction tends to indicate that the decrease in the joint space width is linked to the 236
organization (m-scale) of the vertical trabeculae.
237
M AN US CR IP T
AC CE PT ED
Strengths and limitations 238
This study presents several important strengths. The database used included 319 subjects with 239
publicly available radiographs and grading. Both control knees and incident cases had KL 0 240
at baseline, so our TBT findings are likely representative of early subchondral bone changes 241
that occur before definite radiological OA features are detectable on plain radiographs either 242
in the medial or the lateral compartment. Also, this dataset included multicenter radiographs 243
using different equipment without calibration for TBTp. Neither the variations in the 244
acquisition parameters nor the impact of the soft tissue in the X-ray absorption and diffusion 245
were taken into account. In other words, the descriptors were not adjusted for the equipment 246
or clinical center and the prediction reflected the results of a multicenter cohort without 247
adjustments. Regression models were reduced parsimoniously and cross-validated to avoid 248
highly complex or over-fitted models.
249
Some limitations have also to be considered. The compartments affected with knee OA 250
according to the local TBT were not treated independently and it might be relevant to stratify 251
the ‘cases’ into several groups (e.g. lateral/medial initiators). Moreover, the femoral 252
subchondral bone texture was not assessed although it might be relevant in this task. In the 253
subject inclusion criteria, digitized films were rejected due to the non-standardized protocol 254
used for digitization. The piecewise fractal aspect of the TBT and a fortiori the cut-off scale 255
M AN US CR IP T
AC CE PT ED
multifractal characterization of the TBT might be relevant for early OA diagnosis. Finally, 257
the inclusion of the global TBT was performed using a patchwork of ROIs, yet the 258
interactions between the TBTp were not investigated. This requires further work on the 259
characterization of this global TBT with a more parsimonious and meaningful descriptor than 260
a 64-feature vector.
261
This study highlighted the high selectivity of the actual at-risk patients using the trabecular 262
bone texture. The standard accuracy or AUC values reported in the literature does not reflect 263
details of the sensitivity and specificity of the predictive models. Herein, high values of DOR 264
clearly indicates that TBT analyses can identify a subset of true OA initiators with a very low 265
number of false positives.
266
Conclusion 267
Our findings, along with those reported by Podsiadlo et al. 19, highlight the crucial role of the 268
subchondral bone in the early stage of knee OA. The TBT analysis proposed in this study 269
demonstrated its potential to predict the incidence of radiographic knee OA and highlighted 270
several sub-clusters of texture features (orientations and scales).
271
The quantification of the TBT of the subchondral bone at the tibia needs further investigation 272
to clearly state the links with the actual trabecular microarchitecture. However, it may 273
contribute to the development of a new and feasible imaging biomarker able to predict 274
disease initiation and identify patients at high risk. This predictive ability would be highly 275
M AN US CR IP T
AC CE PT ED
useful in clinical trials to select patients with the highest risk of initiation or progression of 276
knee OA when new disease-modifying treatments become available.
277
Acknowledgments
278
The authors would like to thank the OAI study participants and clinical staff as well as the 279
coordinating center at UCSF. The authors wish to thank the ANR for financial support under 280
the project ANR-12-TECS-0016-01. We acknowledge Mrs Villequenault for typing the 281
manuscript.
282
OAI disclaimer 283
The OAI is a public-private partnership comprising five contracts (N01-AR-2-2258; N01- 284
AR-2-2259; N01-AR-2-2260; N01-AR-2-2261; N01-AR-2-2262) funded by the National 285
Institutes of Health, a branch of the Department of Health and Human Services, and 286
conducted by the OAI Study Investigators. Private funding partners include Merck Research 287
Laboratories; Novartis Pharmaceuticals Corporation, GlaxoSmithKline; and Pfizer, Inc.
288
Private sector funding for the OAI is managed by the Foundation for the National Institutes 289
of Health. This manuscript was prepared using an OAI public use data set and does not 290
necessarily reflect the opinions or views of the OAI investigators, the NIH, or the private 291
funding partners.
292
M AN US CR IP T
AC CE PT ED
Author Contributions 293
All authors contributed to the conception and design of the study, analysis and interpretation 294
of data, drafting the article and revising it critically for important intellectual content, and 295
approved the final version submitted. Dr Lespessailles takes responsibility for the integrity of 296
the work as a whole, from inception to finished article.
297
Role of the funding sources 298
Thomas Janvier (PhD student) was funded by ANR-12-TECS-0016-01.
299
Funding sources had no role in the study design, collection, analysis and interpretation of the 300
data or the decision to submit the manuscript for publication.
301
Conflict of interest 302
All authors state that they have no conflicts of interest.
303
References
304
1. Bitton R. The economic burden of osteoarthritis. Am J Manag Care. 2009;15(8 305
Suppl):S230-S235. doi:10.1002/art.1780290311.
306
2. Losina E, Paltiel AD, Weinstein AM, et al. Lifetime medical costs of knee 307
osteoarthritis management in the United States: Impact of extending indications for 308
total knee arthroplasty. Arthritis Care Res. 2015;67(2):203-215.
309
M AN US CR IP T
AC CE PT ED
doi:10.1002/acr.22412.
310
3. Loeser RF, Goldring SR, Scanzello CR, Goldring MB. Osteoarthritis: A disease of the 311
joint as an organ. Arthritis Rheum. 2012;64(6):1697-1707. doi:10.1002/art.34453.
312
4. Hunter DJ, Felson DT. Osteoarthritis. BMJ. 2006;332(7542):639-642.
313
doi:10.1136/bmj.332.7542.639.
314
5. Kraus VB, Feng S, Wang S, et al. Subchondral bone trabecular integrity predicts and 315
changes concurrently with radiographic and magnetic resonance imaging-determined 316
knee osteoarthritis progression. Arthritis Rheum. 2013;65(7):1812-1821.
317
doi:10.1002/art.37970.
318
6. Bruyere O, Dardenne C, Lejeune E, et al. Subchondral tibial bone mineral density 319
predicts future joint space narrowing at the medial femoro-tibial compartment in 320
patients with knee osteoarthritis. Bone. 2003;32(5):541-545. doi:10.1016/S8756- 321
3282(03)00059-0.
322
7. Roemer FW, Guermazi A. Osteoarthritis Year in Review 2014: Imaging. Osteoarthr 323
Cartil. 2014;22(12):2003-2012. doi:10.1016/j.joca.2014.07.012.
324
8. MacKay JW, Murray PJ, Kasmai B, Johnson G, Donell ST, Toms AP. Subchondral 325
bone in osteoarthritis: Association between MRI texture analysis and 326
histomorphometry. Osteoarthr Cartil. 2016:1-8. doi:10.1016/j.joca.2016.12.011.
327
M AN US CR IP T
AC CE PT ED
knee osteoarthritis progression: data from the Osteoarthritis Initiative. Osteoarthr 329
Cartil. February 2016:45067. doi:10.1016/j.joca.2016.10.005.
330
10. Liebl H, Joseph G, Nevitt MC, et al. Early T2 changes predict onset of radiographic 331
knee osteoarthritis: data from the osteoarthritis initiative. Ann Rheum Dis.
332
2015;74(7):1353-1359. doi:10.1136/annrheumdis-2013-204157.
333
11. Shamir L, Ling SM, Scott W, Hochberg M, Ferrucci L, Goldberg IG. Early detection 334
of radiographic knee osteoarthritis using computer-aided analysis. Osteoarthr Cartil.
335
2009;17(10):1307-1312. doi:10.1016/j.joca.2009.04.010.
336
12. Kwan Tat S, Lajeunesse D, Pelletier JP, Martel-Pelletier J. Targeting subchondral bone 337
for treating osteoarthritis: what is the evidence? Best Pract Res Clin Rheumatol.
338
2010;24(1):51-70. doi:10.1016/j.berh.2009.08.004.
339
13. Burr DB, Gallant M a. Bone remodelling in osteoarthritis. Nat Rev Rheumatol.
340
2012;8(11):665-673. doi:10.1038/nrrheum.2012.130.
341
14. Caldwell CB, Moran EL, Bogoch ER. Fractal dimension as a measure of altered 342
trabecular bone in experimental inflammatory arthritis. J Bone Miner Res.
343
1998;13(6):978-985. doi:10.1359/jbmr.1998.13.6.978.
344
15. Kraus VB, Feng S, Wang S, et al. Trabecular morphometry by fractal signature 345
analysis is a novel marker of osteoarthritis progression. Arthritis Rheum.
346
2009;60(12):3711-3722. doi:10.1002/art.25012.
347
M AN US CR IP T
AC CE PT ED
16. Woloszynski T, Podsiadlo P, Stachowiak GW, Kurzynski M, Lohmander LS, Englund 348
M. Prediction of progression of radiographic knee osteoarthritis using tibial trabecular 349
bone texture. Arthritis Rheum. 2012;64(3):688-695. doi:10.1002/art.33410.
350
17. Messent E a, Ward RJ, Tonkin CJ, Buckland-Wright C. Tibial cancellous bone 351
changes in patients with knee osteoarthritis. A short-term longitudinal study using 352
Fractal Signature Analysis. Osteoarthr Cartil. 2005;13(6):463-470.
353
doi:10.1016/j.joca.2005.01.007.
354
18. Podsiadlo P, Cicuttini FM, Wolski M, Stachowiak GW, Wluka AE. Trabecular bone 355
texture detected by plain radiography is associated with an increased risk of knee 356
replacement in patients with osteoarthritis: A 6 year prospective follow up study.
357
Osteoarthr Cartil. 2014;22(1):71-75. doi:10.1016/j.joca.2013.10.017.
358
19. Podsiadlo P, Nevitt MCC, Wolski M, et al. Baseline trabecular bone and its relation to 359
incident radiographic knee osteoarthritis and increase in joint space narrowing score:
360
directional fractal signature analysis in the MOST study. Osteoarthr Cartil. May 2016.
361
doi:10.1016/j.joca.2016.05.003.
362
20. Woloszynski T, Podsiadlo P, Stachowiak G, Kurzynski M. A dissimilarity-based 363
multiple classifier system for trabecular bone texture in detection and prediction of 364
progression of knee osteoarthritis. Proc Inst Mech Eng H. 2012;226(11):887-894.
365
M AN US CR IP T
AC CE PT ED
21. Nevitt M, Felson D, Lester G. The Osteoarthritis Initiative: Protocol for the Cohort 367
Study.; 2006. http://oai.epi-ucsf.org/datarelease/docs/StudyDesignProtocol.pdf.
368
22. Kothari M, Guermazi A, von Ingersleben G, et al. Fixed-flexion radiography of the 369
knee provides reproducible joint space width measurements in osteoarthritis. Eur 370
Radiol. 2004;14(9):1568-1573. doi:10.1007/s00330-004-2312-6.
371
23. Osteoarthritis Initiative. Central Reading of Knee X-Rays for Kellgren & Lawrence 372
Grade and Individual Radiographic Features of Tibiofemoral Knee OA.; 2016.
373
https://oai.epi-ucsf.org/datarelease/SASDocs/kXR_SQ_BU_descrip.pdf.
374
24. Kellgren JH, Lawrence JS. Radiological Assessment of. Ann Rheum Dis.
375
1957;16(April 2009):494-502. doi:10.1136/ard.16.4.494.
376
25. Altman RD, Gold GE. Atlas of individual radiographic features in osteoarthritis, 377
revised. Osteoarthr Cartil. 2007;15(SUPPL. 1):A1-A56.
378
doi:10.1016/j.joca.2006.11.009.
379
26. Janvier T, Toumi H, Harrar K, Lespessailles E, Jennane R. ROI impact on the 380
characterization of knee osteoarthritis using fractal analysis. In: 2015 International 381
Conference on Image Processing Theory, Tools and Applications (IPTA). Orléans, 382
France: IEEE; 2015:304-308. doi:10.1109/IPTA.2015.7367152.
383
27. Janvier T, Jennane R, Valery A, et al. Subchondral tibial bone texture analysis predicts 384
knee osteoarthritis progression: data from the Osteoarthritis Initiative. Osteoarthr 385
M AN US CR IP T
AC CE PT ED
Cartil. 2017;25(2):259-266. doi:10.1016/j.joca.2016.10.005.
386
28. Messent EA, Buckland-Wright JC, Blake GM. Fractal analysis of trabecular bone in 387
knee osteoarthritis (OA) is a more sensitive marker of disease status than bone mineral 388
density (BMD). Calcif Tissue Int. 2005;76(6):419-425. doi:10.1007/s00223-004-0160- 389
7.
390
29. Wolski M, Podsiadlo P, Stachowiak GW. Directional fractal signature methods for 391
trabecular bone texture in hand radiographs: Data from the Osteoarthritis Initiative.
392
Med Phys. 2014;41(8):81914. doi:10.1118/1.4890101.
393
30. Buckland-Wright C. Subchondral bone changes in hand and knee osteoarthritis 394
detected by radiography. Osteoarthr Cartil. 2004;12(SUPLL.):10-19.
395
doi:10.1016/j.joca.2003.09.007.
396
31. Pothuaud L, Benhamou CL, Porion P, Lespessailles E, Harba R, Levitz P. Fractal 397
Dimension of Trabecular Bone Projection Texture Is Related to Three-Dimensional 398
Microarchitecture. J Bone Miner Res. 2000;15(4):691-699.
399
doi:10.1359/jbmr.2000.15.4.691.
400
32. Jennane R, Harba R, Lemineur G, Bretteil S, Estrade A, Benhamou CL. Estimation of 401
the 3D self-similarity parameter of trabecular bone from its 2D projection. Med Image 402
Anal. 2007;11(1):91-98. doi:10.1016/j.media.2006.11.001.
403
M AN US CR IP T
AC CE PT ED
tomographic microscopy projections. IEEE Trans Med Imaging. 2001;20(5):443-449.
405
doi:10.1109/42.925297.
406
34. Hirvasniemi J, Thevenot J, Kokkonen HT, et al. Correlation of Subchondral Bone 407
Density and Structure from Plain Radiographs with Micro Computed Tomography Ex 408
Vivo. Ann Biomed Eng. 2016;44(5):1698-1709. doi:10.1007/s10439-015-1452-y.
409
35. Lynch JA, Hawkes DJ, Buckland-Wright JC. A robust and accurate method for 410
calculating the fractal signature of texture in macroradiographs of osteoarthritic knees.
411
Med Informatics. 1991;16(2):241-251. doi:10.3109/14639239109012130.
412
36. Winzenrieth R, Michelet F, Hans D. Three-Dimensional (3D) microarchitecture 413
correlations with 2d projection image gray-level variations assessed by trabecular bone 414
score using high-resolution computed tomographic acquisitions: Effects of resolution 415
and noise. J Clin Densitom. 2013;16(3):287-296. doi:10.1016/j.jocd.2012.05.001.
416
37. Bousson V, Bergot C, Sutter B, Levitz P, Cortet B. Trabecular bone score (TBS):
417
available knowledge, clinical relevance, and future prospects. Osteoporos Int.
418
2012;23(5):1489-1501. doi:10.1007/s00198-011-1824-6.
419
38. Harrar K, Hamami L, Lespessailles E, Jennane R. Piecewise Whittle estimator for 420
trabecular bone radiograph characterization. Biomed Signal Process Control.
421
2013;8(6):657-666. doi:10.1016/j.bspc.2013.06.009.
422
39. Perrin E, Harba R, Iribarren I, Jennane R. Piecewise fractional Brownian motion. IEEE 423
M AN US CR IP T
AC CE PT ED
Trans Signal Process. 2005;53(3):1211-1215. doi:10.1109/TSP.2004.842209.
424
40. Glas AS, Lijmer JG, Prins MH, Bonsel GJ, Bossuyt PMM. The diagnostic odds ratio: a 425
single indicator of test performance. J Clin Epidemiol. 2003;56(11):1129-1135.
426
doi:10.1016/S0895-4356(03)00177-X.
427
41. R Core Team. R: A Language and Environment for Statistical Computing. 2013.
428
http://www.r-project.org/.
429
42. Venables WN, Ripley BD. Modern Applied Statistics with S. Fourth. New York:
430
Springer; 2002. http://www.stats.ox.ac.uk/pub/MASS4.
431
43. Kuhn M, Wing J, Weston S, et al. caret: Classification and Regression Training. 2015.
432
http://cran.r-project.org/package=caret.
433
44. Robin X, Turck N, Hainard A, et al. pROC: an open-source package for R and S+ to 434
analyze and compare ROC curves. BMC Bioinformatics. 2011;12:77.
435
45. Liang K-Y, Zeger SL. Longitudinal data analysis using generalized linear models.
436
Biometrika. 1986;73(1):13-22. doi:10.1093/biomet/73.1.13.
437
46. Woloszynski T, Podsiadlo P, Stachowiak GW, Kurzynski M. A signature dissimilarity 438
measure for trabecular bone texture in knee radiographs. Med Phys. 2010;37(5):2030- 439
2042. doi:Doi 10.1118/1.3373522.
440
47. Manek NJ, Hart D, Spector TD, MacGregor AJ. The association of body mass index 441
M AN US CR IP T
AC CE PT ED
influences. Arthritis Rheum. 2003;48(4):1024-1029. doi:10.1002/art.10884.
443
48. Jiang L, Tian W, Wang Y, et al. Body mass index and susceptibility to knee 444
osteoarthritis: a systematic review and meta-analysis. Joint Bone Spine.
445
2012;79(3):291-297. doi:10.1016/j.jbspin.2011.05.015.
446
49. King LK, March L, Anandacoomarasamy A. Obesity & osteoarthritis. Indian J Med 447
Res. 2013;138(AUG):185-193. doi:IndianJMedRes_2013_138_2_185_117544 [pii].
448
50. Rogers MW, Wilder F V. The association of BMI and knee pain among persons with 449
radiographic knee osteoarthritis: A cross-sectional study. BMC Musculoskelet Disord.
450
2008;9(1):163. doi:10.1186/1471-2474-9-163.
451
51. Le Graverand M-PH, Brandt K, Mazzuca SA, Raunig D, Vignon E. Progressive 452
increase in body mass index is not associated with a progressive increase in joint space 453
narrowing in obese women with osteoarthritis of the knee. Ann Rheum Dis.
454
2009;68(11):1734-1738. doi:10.1136/ard.2007.085530.
455
52. Griffith JF, Genant HK. Bone mass and architecture determination: state of the art.
456
Best Pract Res Clin Endocrinol Metab. 2008;22(5):737-764.
457
doi:10.1016/j.beem.2008.07.003.
458
459
M AN US CR IP T
AC CE PT ED
Legends
460
Table 1: Characteristics of the 319 patients – 344 knees included in this study. ∆JSN: the 461
decrease in the joint space width evaluated as an increase in the OARSI JSN grade over 48 462
months (in either the medial or lateral side); ∆TOS: the onset of tibial osteophytes evaluated 463
as an increase in the OARSI TOS grade over 48 months (in either the medial or lateral side);
464
∆KL: the increase in the KL grade with certainty of OA presence after 48 months (KL ≥ 2).
465
Table 2: Details of the significant associations of texture parameters and covariates with 466
knee OA initiation defined using the joint space narrowing (JSN), tibial osteophytes (TOS) 467
and Kellgren-Lawrence (KL) grades. These features were extracted from the models reduced 468
with a stepwise algorithm based on the Akaike parsimony criterion.
469
Table 3: Details of the AUROC and DOR values with 95% confidence interval for 470
radiographic knee OA incidence prediction using the three knee OA initiation definitions 471
based on the joint space narrowing (JSN), tibial osteophytes (TOS) and Kellgren-Lawrence 472
(KL) grades.
473
Figure 1: Flowchart of the initial OAI cohort pruning (n is the number of subjects and k is 474
the number of knees).
475
Figure 2: Knee radiograph mapped with a customized patchwork of ROIs.
476
M AN US CR IP T
AC CE PT ED
Figure 3: TBT variograms computed along the two main loading directions (0° and 90°, 477
respectively the image lines and columns). These curves were averaged for every ROI in the 478
patchwork. The areas µTBT and mTBT correspond to the µ-scale [150 µm; 400 µm] and m- 479
scale [0.6 mm; 2mm].
480
Figure 4: ROC curves obtained for the radiographic knee OA incidence prediction using a 481
10-fold cross validation repeated 300 times. The curves shown are averages of the 300 482
repetitions. The three knee OA initiation definitions based on the joint space narrowing 483
(JSN), tibial osteophytes (TOS) and Kellgren-Lawrence (KL) grades were predicted using the 484
clinical covariates alone and their combination with the trabecular bone texture (TBT) 485
parameters.
486
M AN US CR IP T
AC CE PT ED
Baseline 48 Month controls cases controls cases controls cases
No knees 300 (87.2%) 44 (12.8%) 285 (82.8%) 59 (17.2%) 265 (77%) 79 (23%)
Age (years) 61 (54-69) 65 (59-73) 62 (54-69) 60.5 (54-70) 62 (54-69) 60 (54.5-69.5) 62 (54-69) 61 (55.5 -69) BMI (Kg/m2) 27.7 (24.9-30.5) 27.9 (25.0-30.7) 27.6 (25-30.3) 29.15 (24.8-31.9) 27.3 (24.8-30.3) 30 (25.65-31.3) 27.5 (25.1-30.4) 28.59 (24.6-31.05)
Sex
♀ 146 28 138 36 124 50
♂ 154 16 147 23 141 29
Center
A 57 4 52 9 48 13
C 150 21 147 24 136 35
D 86 17 79 24 75 28
E 7 2 7 2 6 3
JSN* grade
0 344 (100%) 300 (87.2%)
1 0 20 (5.8%)
2 0 20 (5.8%)
3 0 4 (1.2%)
JSN: Joint Space Narrowing
Catergorical values are presented as: no. knees (% of the subset) Age and BMI are presented as: median (1st qu. - 3rd qu.)
ΔKL
ΔJSN ΔTOS
174 (50.6%) Subset
170 (49.5%) 61 (17.8%) 171 (49.7%) 103 (29.9%) 9 (2.6%) 344 (100%)
