Subchondral bone length in knee osteoarthritis: A deep


Download Subchondral bone length in knee osteoarthritis: A deep


Preview text

medRxiv preprint doi: https://doi.org/10.1101/2021.04.28.21256271; this version posted May 1, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC-ND 4.0 International license .
Subchondral bone length in knee osteoarthritis: A deep learning derived imaging measure and its association with
radiographic and clinical outcomes
Gary H. Chang1, , Lisa K. Park1, , Nina A. Le1, , Ray S. Jhun1, Tejus Surendran1, Joseph Lai1, Hojoon Seo1, Nuwapa Promchotichai1, Grace Yoon1, Jonathan Scalera2, Terence D. Capellini3,
David T. Felson4, §, Vijaya B. Kolachalama1, 5, §, *
1Section of Computational Biomedicine, Department of Medicine, Boston University School of Medicine, Boston, MA, USA – 02118
2Department of Radiology, Boston University School of Medicine, Boston, MA, USA – 02118 3Department of Human Evolutionary Biology, Harvard University, Cambridge, MA, USA – 02138; Broad Institute of MIT and Harvard, Cambridge, MA, USA 02142 4Section of Rheumatology, Department of Medicine, Boston University School of Medicine,
Boston, MA, USA – 02118; Centre for Epidemiology, University of Manchester and the NIHR Manchester BRC, Manchester University, NHS Trust, Manchester, UK
5Department of Computer Science and Faculty of Computing & Data Sciences, Boston University, Boston, MA, USA – 02215
Joint first authors §Joint senior authors
*Corresponding author: Vijaya B. Kolachalama, PhD Boston University School of Medicine 72 E. Concord Street, Evans 636, Boston, MA, USA – 02118 Email: [email protected] Phone: 617-358-7253
NOTE: This preprint reports new research that has not been certified by peer review and should not be used to guide clinical practice.

medRxiv preprint doi: https://doi.org/10.1101/2021.04.28.21256271; this version posted May 1, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC-ND 4.0 International license .
ABSTRACT
Objective: Develop a bone shape measure that reflects the extent of cartilage loss and bone flattening in knee osteoarthritis (OA) and test it against estimates of disease severity.
Methods: A fast region-based convolutional neural network was trained to crop the knee joints in sagittal dual-echo steady state MRI sequences obtained from the Osteoarthritis Initiative (OAI). Publicly available annotations of the cartilage and menisci were used as references to annotate the tibia and the femur in 61 knees. Another deep neural network (U-Net) was developed to learn these annotations. Model predictions were compared with radiologist-driven annotations on an independent test set (27 knees). The U-Net was applied to automatically extract the knee joint structures on the larger OAI dataset (9,434 knees). We defined subchondral bone length (SBL), a novel shape measure characterizing the extent of overlying cartilage and bone flattening, and examined its relationship with radiographic joint space narrowing (JSN), concurrent WOMAC pain and disability as well as subsequent partial or total knee replacement (KR). Odds ratios for each outcome were estimated using relative changes in SBL on the OAI dataset into quartiles.
Result: Mean SBL values for knees with JSN were consistently different from knees without JSN. Greater changes of SBL from baseline were associated with greater pain and disability. For knees with medial or lateral JSN, the odds ratios between lowest and highest quartiles corresponding to SBL changes for future KR were 5.68 (95% CI:[3.90,8.27]) and 7.19 (95% CI:[3.71,13.95]), respectively.
Conclusion: SBL quantified OA status based on JSN severity. It has promise as an imaging marker in predicting clinical and structural OA outcomes.
Keywords: Knee Osteoarthritis, Joint Space Narrowing, Knee Replacement, Magnetic Resonance Imaging, Deep Learning

medRxiv preprint doi: https://doi.org/10.1101/2021.04.28.21256271; this version posted May 1, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC-ND 4.0 International license .
INTRODUCTION
Knee osteoarthritis (OA) is one of the most common debilitating conditions among older adults [1], with a burden that is expected to increase around the globe due to factors such as increasing rates of obesity [2]. Because there is no effective disease modifying therapy for knee OA, current clinical management relies in part on identification of radiographic abnormalities such as joint space narrowing (JSN) to assist in diagnosing OA and/or estimate likelihood of progression to severe OA [3-6]. Since it is important to accurately evaluate OA pathophysiology, there is a need to develop additional imaging markers, and in some cases, improve the measurement sensitivity and specificity of existing imaging markers to better predict disease progression and potential clinical outcomes.
Magnetic resonance imaging (MRI) can allow us to visually differentiate joint tissues, which facilitates accurate characterization of several imaging markers of knee OA [7]. Previously, a case-control study on a subset of data from the Osteoarthritis Initiative (OAI) showed that bone shape derived from knee MR images predicted later onset of radiographic knee OA [8]. In other studies, researchers used MRI to identify increased tibial plateau size and subchondral bone attrition during pre-radiographic OA stage [9], and also found variability in knee articular surface geometry in individuals with and without OA [10]. Most of these studies were performed on a small subset of individuals. While findings from these important investigations established a proof-of-principle, it is not trivial to extend such studies to a larger cohort due to the sheer volume of cases and the manual labor needed to precisely annotate different anatomic structures in the knee. Recently, Bowes and colleagues proposed a machine learning-driven measure of femur bone shape (defined as B-score), on a large set of individuals from the OAI cohort (n=4,796) [11]. They showed that B-score was associated with risk of current and future pain, functional limitation and total knee replacement. Few other studies have also attempted to characterize bone shapes in OA [8, 12]. The major contributor to the bone shape metric has been the crust of osteophytes that arises at the margin of the cartilage plate in advanced OA. These marginal osteophytes are probably not a source of pain [13], and their size and number are poor proxies for the severity of nearby cartilage loss [14], which is the signature pathologic feature of OA. We developed a bone shape measure defined as subchondral bone length (SBL), which did

medRxiv preprint doi: https://doi.org/10.1101/2021.04.28.21256271; this version posted May 1, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC-ND 4.0 International license .
not include these marginal osteophytes, but rather was driven by the extent of overlying cartilage, decreasing with cartilage denudation. Bone also flattens with increasing disease severity, increasing its area [15, 16]. We hypothesized that SBL variations increase with the severity of radiographic knee OA, as it accounts for the dynamic changes that occur within the subchondral region due to cartilage loss and bone flattening. We then tested whether SBL correlated with severity of knee OA defined by JSN, concurrent knee pain and disability as well as subsequent partial or total knee replacement.

medRxiv preprint doi: https://doi.org/10.1101/2021.04.28.21256271; this version posted May 1, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC-ND 4.0 International license .
METHODS
Study population Data were obtained from the Osteoarthritis Initiative (OAI), an NIH funded observational study of persons with or at-risk of knee OA. Study subjects included men and women ranging in age from 45-79 with or at-risk for symptomatic tibiofemoral knee OA (Table 1). Minorities made up 20.9% of the population of whom 87.3% were African Americans, 4.5% Asians, and 8.2% fell into the “Other” category. Subjects with contraindications for 3T MRI (such as inflammatory arthritis) were excluded from the OAI study (Table 1). The majority of knees at baseline had radiographic severity of KL grades 0-3, with KL grade 4 making up ~3% of the population. Note that knees with information regarding KL grade also had data regarding JSN grade. All the cases which had both KL and JSN grades were included in the study. The severity of JSN was based on centrally read scores ranging from 0 to 3 (0=normal, 1=mild, 2=moderate and 3=severe) by compartments [17].
MRI acquisition and measurements Three-dimensional (3D) dual-echo in steady state (DESS) sagittal MR sequences of the left and right knees were available on the OAI dataset at baseline (n=9,434 knees). All the scans along with the subject-level baseline clinical data are available on the NIH website (https://nda.nih.gov/oai/), and will be forwarded upon request. We also obtained detailed segmentation masks of the femur cartilage, lateral tibia cartilage and medial tibia cartilage for a small portion of knees from the OAI database (n=88), which were used to train and validate the model. The cartilage and the meniscal image masks were provided by the OA Biomarkers Consortium FNIH Project and performed by iMorphics (Manchester, UK). As per the OAI documentation (https://nda.nih.gov/oai/study_documentation.html), the knees were chosen to represent the OAI database population (primarily moderate and severe KL grades, Male = 45, Female = 43) and should be applicable for the validation process of our study.
The DESS sequence images provide detailed definition of 3D structures and their shapes, particularly of cartilage morphology [18, 19]. The imaging was performed with a 3.0 Tesla magnet using imaging sequence with TR/TE of 16.3/4.7 ms. We selected knees (n=4,727 right

medRxiv preprint doi: https://doi.org/10.1101/2021.04.28.21256271; this version posted May 1, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC-ND 4.0 International license .
and 4,707 left knees) from the OAI baseline exam with DESS sequence MR images. The DESS sequence provides high in-plane and out-of-plane resolution (0.7 mm × 0.365 mm × 0.456 mm) in a time efficient manner. The images encompassed the cartilage of the knee joint as well as the subchondral structures of the tibia and the femur.
Image preprocessing and a fast region-based convolutional neural network Image registration and quality check were performed on all the knees as previously described [20]. To focus the network on the joint area, a large-sized bounding box with a dimension of 272x240 pixels was created to identify regions of interest (ROI) comprised of all the subchondral compartments containing the manually annotated femoral and tibia cartilages. The coordinates of the ROI of each knee were subsequently used to train a fast region-based convolutional neural network (Fast R-CNN) model [21] to automatically detect the region of the knee joint (Figure S1). The sagittal MRI slices subsequently underwent histogram equalization and their intensity values were normalized to a range of 0 to 1. These slices were then used as the inputs to the neural network used to extract the knee joint structures. More details are available in the supplement.
Image annotation pipeline From the OAI database, 88 knee MRIs had expert-driven annotations of the cartilage. These MRIs were chosen as our primary dataset for training and validating the model for both automated segmentation of cartilage and bone shape (Figure S2). The 88 knee MRIs were randomly split in the ratio of 7:3 for training and validation, respectively. The 61 knee images for training (n=1,753 2D sagittal slices) were passed through an image-processing pipeline based on Distance Regularized Level Set Evolution (DRLSE) to extract the bone shapes (Figure S3). DRLSE is an edge-based active contour model that uses the gradient information of the images to expand the segmented area until it meets the boundary [22]. Once the active contour model was applied, the obtained bone shapes were further manually verified and adjusted to exclude erroneously captured soft tissues in the bone shapes. Finally, the expert-driven segmented regions of cartilage and meniscal areas were superimposed to adjust the bone shapes, where needed. The modified bone shapes from the 61 cases were then used to train a deep neural network (See below and the supplement). The remaining cases (794 2D sagittal slices from 27

medRxiv preprint doi: https://doi.org/10.1101/2021.04.28.21256271; this version posted May 1, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC-ND 4.0 International license .
knees) were used for model validation and underwent expert-assisted manual annotation of the bone shapes as detailed below.
Annotation of the knee structures A board-certified radiologist (JS) manually annotated bone shapes of the knee on the 27 test cases. Using a stylus on a touch screen device, the expert outlined the cortical surface of the tibia and the femur. We traced the SBLs on both the tibia and the femur (direction shown in dotted yellow lines in Figure 1A), to capture the shape of the cortical bone of the femur or the tibia in contact with the cartilage.
Deep learning framework for bone and cartilage shape segmentation All the cases reviewed by the expert radiologist were used to train a deep neural network to simultaneously extract the bone and cartilage shapes of the tibia and the femur (Figure S4). The neural network was based on a well-known deep learning architecture (U-Net) [23], and has the capability to learn different patterns within imaging data. More details can be found in the supplement.
Measurement of subchondral bone length We extracted the SBLs on both the tibia and the femur to capture the shape of the cortical bone of the femur or the tibia that is in contact with the cartilage. Briefly, we applied a distance transform on the output of the U-Net model to detect the edge of the bone region in contact with the cartilage (Figure 1B). The distance transform measures the distance of each pixel on the bone from its closest pixel on the cartilage. We subsequently skeletonized the detected region to a thickness of one pixel and calculated its arc length. If there were multiple regions detected in the femur or the tibia for a specific 2D sagittal slice, then we skeletonized each region individually and summed the arc lengths of each segment as SBL. For instance, areas denuded of cartilage would produce a gap in the bone length, thus the computed SBL would be the sum of the segments of the cartilage-covered bone. These measurements were confirmed by the radiologist.

medRxiv preprint doi: https://doi.org/10.1101/2021.04.28.21256271; this version posted May 1, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC-ND 4.0 International license .
The SBL estimates for both the femur (red curved lines in Figures 1C & 1D), and the tibia (blue curved lines in Figures 1C & 1D), were extracted from each 2D sagittal slice at various locations along the frontal axis (medial to lateral side) of each knee (Figures 2A, 3A & 4A). The knees were further stratified by JSN grade within the medial and lateral compartments on the femur and the tibia.
Statistical analysis The deep neural network used for segmenting the bone shapes was evaluated by computing the Dice coefficient between the predicted masks (generated from the U-Net model), and the bone shapes outlined by the radiologist on the test cases (n=27). Descriptive statistics are presented as mean values and standard deviations. A Student’s t-test was used to compare the mean value of two different groups. A p-value of <0.05 was considered statistically significant. To compare SBLs between knees of various sizes, we took the calculated SBL values from the MRI slices where cartilage was visible and interpolated the values uniformly along the frontal axis (Figures 2A, 3A & 4A). For the SBLs taken from the nth location along the frontal axis, we applied a ttest comparing the SBLs of knees with only lateral JSN (374 right and 342 left knees) with knees with no JSN (5128 right and left knees), and knees with only medial JSN (1604 right and 1504 left knees) with knees with no JSN (5128 right and left knees).
We computed mean estimates of femur-specific and tibia-specific SBL measurements at each location on the frontal axis for all the knees with JSN=0. For each knee with JSN>0, we computed difference between the SBL value measured at the nth location and the corresponding mean SBL estimate at that location in knees with JSN=0. We then added the absolute value of these differences to create knee-specific SBL measures and divided them into quartiles. Each group (medial or lateral JSN cases) was further stratified based on clinical outcomes, which included baseline WOMAC pain and disability scores and future total knee replacement. The
baseline pain and disability scores were divided by severity, such that scores for pain ≥4 and <8 were moderate and those ≥8 were severe. Similarly, scores for disability ≥20 and <35 were moderate and those ≥35 were severe [11]. The criterion for future knee replacement was the
subject having a knee replacement seen on follow-up x-ray at any time in an 8-year follow-up after baseline.

medRxiv preprint doi: https://doi.org/10.1101/2021.04.28.21256271; this version posted May 1, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC-ND 4.0 International license .
We calculated odds ratios by comparing the odds of each quartile developing the various outcomes, using Quartile 1 as a reference. The odds ratios were calculated using a 2×2 contingency table with binomial outcomes for the pain and disability scores such that one outcome was compared to the remaining outcomes in its corresponding category. For example, the odds of having moderate pain were compared to no pain and severe pain combined. Severe pain was compared with the combination of moderate and no pain. The p-values were calculated using the Fisher exact test. Python scripts are made available on GitHub (https://github.com/vkola-lab/ar2021).

medRxiv preprint doi: https://doi.org/10.1101/2021.04.28.21256271; this version posted May 1, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC-ND 4.0 International license .
RESULTS
Manual versus automated segmentation Using the deep learning framework, we estimated the bone shapes defining the tibiofemoral joint (Figure S5). The Dice coefficients for the femur cartilage, medial and lateral tibia cartilage, on the U-Net model were 0.903, 0.913, and 0.873, respectively. Comparison of the manual annotation and automated segmentation of the femur, the tibia, and the meniscus with Dice coefficients demonstrated the similarity between the predicted and expert-driven segmentations for the U-Net model. The red and blue lines in the figures indicate SBLs connected with the femoral and tibial cartilage, respectively (Figures 1C & 1D). In a further comparison between the respective masks generated by the U-Net model (Predicted) and the expert-driven hand annotated images, scatter plots were generated with the measured SBL of each MRI slice of the reference estimate against predicted values (n=27 subjects). Reference estimate versus predicted SBLs of femur generated an R2 value of 0.922 (Figure S6A). Much of the discrepancy between the reference estimate from the radiologist versus predicted SBL values occurred at the extreme medial and lateral sagittal slices (SBL length<50 pixels). This pattern was consistent with the findings in tibial SBL (R2=0.902) (Figure S6B), with higher discrepancies seen at SBLs of less than 25 pixels. These correspond to the lateral-most and medial-most slices of the knees.
Subchondral bone length in relation to joint space narrowing Our analysis showed that most SBL values from knees with JSN >0 were significantly different in length than knees with no JSN for both the femur and the tibia (Figures 2B-2D & 3B-3D). This was true for knees with both medial and lateral JSN. In cases of medial JSN, on the femur we observed significant differences in the SBL values for most of the medial-central, central and lateral-most regions (Figures 2C). In the cases of lateral JSN, the SBL values were significantly different from JSN=0 knees in most of the lateral regions of the femur (Figure 2D), except for a few lateral-most regions for JSN=3. For cases with lateral JSN, statistically significant differences were not observed in most of the medial-central regions of the femur, regardless of the JSN grade (Figure 2D).

Preparing to load PDF file. please wait...

0 of 0
100%
Subchondral bone length in knee osteoarthritis: A deep