Detection of bone loss via subchondral bone analysis

Abstract

To date, there is no single sign, symptom, or test that can clearly diagnose early stages of Temporomandibular Joint Osteoarthritis (TMJ OA). However, it has been observed that changes in the bone occur in early stages of this disease, involving structural changes both in the texture and morphometry of the bone marrow and the subchondral cortical plate. In this paper we present a tool to detect and highlight subtle variations in subchondral bone structure obtained from high resolution Cone Beam Computed Tomography (hr-CBCT) in order to help with detecting early TMJ OA. The proposed tool was developed in ITK and 3DSlicer and it has been disseminated as open-source software tools. We have validated both our texture analysis and morphometry analysis biomarkers for detection of TMJ OA comparing hr-CBCT to μCT. Our initial statistical results using the multidimensional features computed with our tool indicate that it is possible to classify areas of demonstrated loss of trabecular bone in both μCT and hr-CBCT. This paper describes the first steps to alleviate the current inability of radiological changes to diagnose TMJ OA before morphological changes are too advanced by quantifying subchondral bone biomarkers. This paper indicates that texture based and morphometry based biomarkers have the potential to identify OA patients at risk for further bone destruction.

1. INTRODUCTION

To date, there is no single sign, symptom, or test that can clearly diagnose early stages of Temporomandibular Joint Osteoarthritis (TMJ OA). However, it has been observed that changes in the bone occur in early stages of this disease, involving microstructural changes to the bone marrow and the subchondral cortical plate. The subchondral bone in the TMJ condyle is the site of numerous dynamic morphological transformations, which are part of the initiation/progression of OA, not merely secondary manifestations to cartilage degradation. Current measurement of bone marrow architecture are limited or subjective, such as subchondral bone volume and density.[1,2] Volumetric changes cannot detect structural changes at specific locations alone, since they lack the ability to quantify subtle gray-level variations within the low-density bone 3D-structure. Subchondral trabecular bone architecture is highly anisotropic, and the mechanical properties of the tissue are therefore different in the different planes.

Texture is an intuitive concept heuristically but difficult to define precisely. It can be defined as series of homogeneous visual patterns that are observed in certain kinds of materials. Humans describe texture through qualitative concepts such as fine, coarse, granulated or smooth. These descriptions are not precise and, in addition, they are not quantitative. Texture quantification has been studied for a long time and many different texture quantification and analysis techniques exist[3–7]. Morphometry refers to the quantitative analysis of form. Histomorphometry was the first type of bone morphometric analysis performed, and it consists in slicing pieces of ex-vivo bone and performing a succession of 2D morphometric analyses on the tissue slices obtained. This technique was limited by the destructive nature of the procedure, which did not allow application to in-vivo patients. Additionally, due to the 2D nature of the images, certain types of features such as bone volume density (Bv/Tv) and bone surface density (Bs/Bv)[8] could be computed, but the computation of other types of 3D features, such as trabecular thickness (Tb.Th), trabecular separation (Tb.Sp), and trabecular number (Tb.N), was not possible[9].

Subchondral bone texture and morphometry measured in the 3D space can potentially provide an automatic and objective measure of subchondral bone architecture using high-resolution Cone Beam CT[10] (hr-CBCT).

In this paper we present a tool to detect and highlight subtle variations in subchondral bone texture and morphometry combined, which are obtained from hr-CBCT in order to help with detecting early TMJ OA. We have validated the texture imaging biomarkers for detection of TMJ OA (bone sclerosis) comparing hr-CBCT to μCT. Areas displaying lacunae in the bone trabecular texture were verified using histology.

2. MATERIALS

We have used 16 condylar bone specimens, obtained from 12 patients with diagnosis of TMJ OA who were treated with surgical resection. From those 16 condylar bone specimens, we have 4 pairs of specimens that come from the same patient (left and right condyles). The specimens were originally preserved in 70% ethanol at freezing temperatures. Imaging of all specimens was acquired using μCT and hr-CBCT. For μCT [11] image acquisition (μCT Scanco Medical, Bassersdorf, Switzerland, protocol settings were: voxel size 34.4 μm, 90 kVp, 155 μA, medium resolution, 0.5 mm AL filter, and integration time 500 ms), specimens were first embedded in 1% agarose, placed in a 34 mm diameter tube and scanned. The proposed hr-CBCT protocol is a TMJ specific protocol designed for this study, that has a reduced field-of-view for better signal to noise ratio (SNR). hr-CBCT were acquired using 3D Accuitomo13, JMorita, with the parameters FOV: 40 × 40 mm, high resolution scanning mode, 90 kV, 5 mA, 30.8 s, 125 μm voxel size, effective dose 114 μSv.

We use histological analysis (see “Methods, Histology methods”) to confirm the presence of bone loss and then generated 35 cropped images (sub-images) from the original CBCT and μCT images, obtained from areas that displayed lacunae in the subchondral bone trabecula (bone loss) in histological analysis (we called these cubes “defects” D) as well as unaffected bone (we call these cubes are called “trabecula” T). See figure 1 for examples of our input data.

Figure 1.

Example of input data. TMJ condyle acquired with a) μCT and d) hr-CBCT. Healthy trabecular space acquired with b) μCT and e) hr-CBCT. Trabecular space with bone sclerosis acquired with c) μCT and f) hr-CBCT.

3. METHODS

3.1 Histology methods

We performed histology to verify the presence of subchondral bone alterations in different patches of subchondral bone. The bone samples were decalcified in 10% EDTA solution for 6 weeks. After decalcification, the specimens were dehydrated in a graded series of ethanol and embedded in paraffin. Serial histological sections 6 μm thick were cut from the areas of interest, which were identified using hr-CBCT. Measurements from one of the outer surfaces (surface of reference) of a specific bony sample was used for identification of regions of interest. Areas suggestive of lacunae defects (subchondral bone alterations) in each sample were identified using ITKSnap[12], by counting the number of slices from the surface of reference used for paraffin inclusion. Hematoxylin and eosin (WAKO, Osaka, Japan) histology of those sections provided an overview of the presence of osteoclastic lacunae. Sections from the portion of the region of trabecula defect as evaluated by hr-CBCT images were evaluated with TRAP staining to enumerate osteoclast numbers. After washing in 0.1 M acetate buffer (pH 5.0), histological sections were incubated with a mixture of naphtol ASMX phosphate as substrate and red violet LB salt (Sigma, St Louis, Mo) diluted in 0.1 M acetate buffer (pH 5.0) containing 50 mML tartaric acid at 37° C for about 30 minutes. Osteoclasts were defined as TRAP-positive multi-nucleated cells on the bone surface. Bone surfaces were measured using a computer-assisted histomorphometric analyzing system (Osteomeasure; OsteoMetrics Inc. Atlanta, GA) as previously described[13] and using standardized protocol and nomenclature[14]. Presence of osteoclasts was confirmed by 2 observers for each specimen for the extraction of image sub-patch generation. Areas that presented results of osteoclast presence and visual lacunae areas in the trabecular spaces of our bone samples were sub-sampled from the image as “bone defects” for further analysis.

3.2 Texture analysis methods

We have investigated two different set of textural features that are sensitive to different aspects of bone architecture: co-occurrence based-features[5] (concentrated in local neighborhoods and connectivities) and run-length based-features[6] (based in grey-level clusters). The computation of co-occurrence features (or Haralick features) is based on the grey level co-occurrence matrix (GLCM) computed voxel’s 3D neighborhood. The GLCM matrix describes intensity organization of each voxel’s neighborhood thanks to its second-order joint probability function. The computation of the run length features is based on the grey level run length matrix (GLRLM) computed also for each voxel’s neighborhood. A grey-level run is a set of consecutive, collinear picture points having the same grey-level value. The length of the run is the number of picture points in the run. Both the GLCM and GLRLM matrices describe each neighborhood local texture.

In our specific implementation, we computed 18 textural features from these matrices: 8 co-occurrence features, and 10 run length features. These features are: energy, entropy, correlation, Inverse Difference Moment (IDM), contrast, cluster shade, cluster prominence, Haralick’s correlation, short run emphasis (SRE), long run emphasis (LRE), grey level non-uniformity (GLN), run length non-uniformity (RLN), low grey level run emphasis (LGRE), high grey level run emphasis (HGRE), short run low grey level emphasis (SRLGE), short run high grey level emphasis (SRHGE), long run low grey level emphasis (LRLGE), and long run high grey level emphasis (LRHGE)[4].

3.3 Morphometry analysis methods

We have investigated five different features that capture different features of the 3D shape of subchondral bone trabecular structure: bone volume density (Bv/Tv), bone surface density (Bs/Bv)[8], trabecular thickness (Tb.Th), trabecular separation (Tb.Sp), and trabecular number (Tb.N).

Bone volume density or BvTv indicates the fraction of a given volume of interest (Total Volume Tv) that is occupied by mineralized bone (Bone Volume Bv). Bone surface density or BsBv represents how many bone lining cells cover a given volume of bone (Bv). Trabecular number (TbN) is taken as the inverse of the mean distance between the mid-axes of the structure to be examined. Trabecular thickness (TbTh) is determined by filling maximal spheres into the structure using a distance transform. Then the average thickness of all maximal spheres is calculated to give an estimate of mean TbTh. Trabecular separation (TbSp) is calculated in the same way than TbTh, but this time the voxels representing non-bone parts are filled with maximal spheres. TbSp can thus be expressed as the average thickness of the marrow cavities.

These features in based in the following parameters, N_total. N_bone and N_boundary (see figure 3). N_total is the total number of voxels in the volume of interest, Nbone represents the number of voxels in the bone area and N_boundary is the number of voxels that are part of the bone/non-bone boundary in the volume of interest. N_boundary can be separated for each direction (N_boundaryX, N_boundaryY and N_boundaryZ).

Figure 3.

(left) NTotal, (center) NBone, highlighted in marron, and (right) NBoundary highlighted in marron lines.

3.4 Open-source bone analysis tools

We have implemented both these texture and morphometry analysis filters as part of the Insight Toolkit (ITK)[15]. ITK is an open-source, cross-platform system that provides developers with an extensive suite of software tools for image analysis. Because ITK is an open-source project, developers from around the world can use, debug, maintain, and extend the software. This decision was made to ensure the work and funds invested in this effort will survive beyond the scope of this project, and increase its potential to be of help to more research programs.

One of the most important features of this implementation[16] is that these filters improve computational cost because they are multi-threaded: they take advantage of ITK’s fast neighborhood operators, and they re-use intermediate computations and minimize memory use. We are proud to report that this state of the art and first of its kind implementation of texture features takes only 2 minutes per high-resolution 3D medical imaging volume, instead of several hours. Our implementation of morphometry features is also very computationally efficient.

Also, these filters are the only implementation we are aware of that can compute 3D maps of multidimensional features (N-D, N = number of dimensions). These maps are comprised of a collection of features (see section “Texture analysis methods” and “Morphometry analysis methods” above), which are computed per voxel. Additionally, we have also implemented python wheel packages that allow to easy installation of the ITK texture filters (see equations 1 and 2) and all its dependencies in order to have them ready to use in python code.

$$\text{python}\text{-m}\text{pip}\text{install}\text{itk-texturefeatures}$$ (1)

$$\text{python}\text{-m}\text{pip}\text{install}\text{itk-bonemorphometry}$$ (2)

They have been generated for three main operating systems: Mac, Linux and Windows, and three versions of python: 2.7, 3.5 and 3.6. We believe python wheels will ensure that other research labs will be able to adopt this technology for their own custom biomedical applications.

3.5 Statistical analysis

The goal of the experiments below are to test the discriminant properties of both μCT and CBCT data for different types of trabecular spaces (healthy vs those presenting bone degradation verified by histology), as well as the relationship between μCT as our gold standard and the TMJ specific hr-CBCT protocol. For that we used Principal component analysis (PCA) and Distance Weighted Discrimination (DWD)[17], and validate the results with direction-projection-permutation (DiProPerm) [18] hypothesis tests.

PCA is a statistical procedure that uses an orthogonal transformation to convert a set of observations of possibly correlated variables into a set of values of linearly uncorrelated variables called principal components (or sometimes, principal modes of variation).

DWD [17] created a model that helps differentiate our T/D textural+morphometry feature samples. DWD was developed as an alternative binary linear classifier to support vector machines (SVM), and it performs very well compared with competing methods. The goal is to take information from every data point into account when determining the separating hyperplane rather than just a small number of support vectors. The hyperplane, determined by a vector w and scalar β, is found by solving a second-order cone program of the form shown in equation 3 (describing the DWD optimization problem formulation), where each x_i is a d-dimensional observation, and each y_i is the class label of x_i (+1 or −1).

$$\underset{r,w,\beta ,\xi }{\mathit{min}}{\sum }_{i=1}^n\left(\frac{1}{r_i}+C{\xi }_i\right);0\le r_i=y_i(x_i^{\prime }w+\beta )+{\xi }_i,\xi \ge 0,{\Vert w\Vert }^2\le 1$$ (3)

The r_i are meant to represent the residuals from each data point to the hyperplane, so by minimizing the sum of reciprocals we maximize their distance to the hyperplane with more weight placed on points closer to the boundary. ξ is an error vector, penalized by a constant C in the objective, which allows for some data points to fall on the wrong side of the hyperplane if necessary[17].

DiProPerm[18] is a permutation-based hypothesis test that assesses the chance that the observed degree of separation happened as a result of expected random variation. It was developed with DWD in mind as an area of application, but it represents a general framework of nonparametric hypothesis testing built to discern typical and atypical behavior in high-dimensional settings. To conduct the test, we randomly shuffle the class data labels and project it onto the DWD direction determined by the two new classes. We record the mean difference and repeat the process 1000 times to create an empirical distribution of mean differences. The p-value of the test is the percent of the empirical distribution larger than the mean difference between the original classes.

4. RESULTS

We used the entire array of 18 bone textural features except for 6 that did not display any individual discriminatory behavior, as well as the 5 bone morphometry features (17D) and we look at the variability and separability of such vectors obtained from our 35 sub-images in the PCA and DWD spaces for both μCT and hr-CBCT.

The main diagonal of plots in the grid in figure 3a displays the projections of the μCT 17D samples in each PC, with corresponding pairwise scatterplots off the diagonal. We see reasonable separation of classes (unaffected trabecula T = blue and bone defects D = red) in most of the orthonormal basis combinations, especially in the PC1&2 projections. Performing the same analysis with the 17D samples computed from hr-CBCT, we still see a reasonable separation of classes, perhaps a bit worse than in μCT (see figure 3b). We see reasonable separation of classes (T = blue and D = red) in most of the orthonormal basis combinations, especially in the PC2&3 projections.

We also calculated the direction that best separates our two classes of 17D textural features using DWD. The resulting DWD projection and first three orthogonal PC directions are displayed in figure 4 for both μCT and hr-CBCT. For the μCT results displayed above we see how DWD achieves good separation of classes (figure 4a). The area under the curve (AUC) obtained from the region receiving operating characteristic (ROC) curve analysis computed from the 17D textural features demonstrates very good separability value (0.853). This demonstrates these features work better together than alone when it comes to detecting bone quality. This good separability is maintained for samples computed from hr-CBCT, as displayed by the DWD/PCA plots in figure 4b. The AUC computed from the ROC analysis for the 17D textural samples computed from hr-CBCT displays a slightly better value (0.858).

Figure 4.

PCA analysis of 17D texture & morphometry features from cropped sub images in a) μCT and b) hr-CBCT

The DiProPerm hypothesis test for this data (see figure 5) in μCT and hr-CBCT provide p-values of 0.07 and 0.006, respectively. By this metric the separation between classes is quite significant in hr-CBCT but not in μCT. This might indicate CBCT reconstruction-induced smoothing removes some of the apparent noise existing in the μCT data.

Figure 5.

DWD analysis of 17D vectors computed from cropped sub images in a) μCT b) hr-CBCT.

5. DISCUSSION

Subchondral trabecular bone changes such as metabolism and sclerosis, as well as modification of trabecular architecture happen before symptoms occur[19,20]. The inability to quantify phenotypes of abnormal subchondral bone texture is a severe bottleneck in understanding early stages of OA. The work presented in this paper contains the first steps to alleviate this information bottleneck by quantifying subchondral bone biomarkers that have the potential to identify OA patients at risk for further bone destruction.

Our initial results using the multidimensional features computed with our ITKTextureFeatures[21] and ITKBoneMorphometry[22] indicate that it is possible to classify areas of demonstrated trabecular bone loss bone sclerosis in both μCT and hr-CBCT. Both of those tools are being disseminated as open-source free software tools. Commercial software packages, such as Geomagic,[23] InVivo,[24] and Amira,[25] produce excellent surface reconstructions, but they are not open source, cannot be modified, and do not interact well with each other. More importantly, they have no support for tracking subchondral bone remodeling nor do they provide patient specific solutions for OA. In addition, commercial software often locks researchers to a single vendor and hinders new clinical discoveries.

To detect early variations in health and disease of the TMJ condyle OA bone architecture, the structure of the subchondral trabecular bone in-vivo using hr-CBCT has yet to be investigated. Future steps will look into calculating bone texture phenotypes using the proposed technology in scans obtained from in-vivo patients.

Figure 2.

a) full condyle bony sample b) hr-CBCT and μCT scans c) histology results, arrows indicate presence of osteoclasts

Figure 6.

DiProPerm results of the 17D vectors DWD classification computed from cropped sub images in a) μCT b) hr-CBCT.